# Adeel Ahmad

Last updated October 2026

I'm a software engineer in Melbourne. This page is for people who have my CV and want to understand how the different parts of it connect.

## How it fits together

### Grade 8 to 2006, Lahore — Starting with security

I started poking at networks in grade 8. In 2006, still at school, I did my first formal vulnerability assessment, on a company's website, and had to explain what I found to the people who ran it.

### 2007 to 2012 — Small companies, every job

I studied computer science while working on .NET and PHP projects. The companies were small, so the developer also looked after servers, the network and the phone system. That is where VMware, Asterisk and Cisco came in.

### 2013 to 2022 — Running a consultancy

I ran Xoho Tech for nine years and it grew to about 40 people. It stayed small enough that I kept writing code. Power cuts kept wiping out developers' work, so I moved our development into the browser. That led to containers, a home lab to understand virtualisation and storage, and later Firecracker microVMs. Our archive and media projects needed search, OCR and speech recognition, which was my first contact with machine learning.

### 2021 to 2023 — Moving to cloud, starting with AI

I started experimenting with OpenAI's API in 2021. In 2022 I moved to Melbourne to work as an AWS consultant. In 2023 I sketched how an AI could prepare a plan and first changes before a developer picks up a task, built a small AutoGen prototype, and tried running small models on my own hardware.

### 2024 — Sensitive data

I kept experimenting with fine-tuning and small models at home, and gave a public talk on building a personal "second brain" on AWS serverless, using small models for the cheap steps and a large one only when needed. At work I became tech lead on a clinical AI product, where every statement in a generated note had to point back to the conversation it came from.

### 2025 to 2026 — AI at a bank, models at home

At Commonwealth Bank I work on AI applications from early design to production, mostly on identity, access and security. At home I train small models on a Mac to understand how they learn to reason, and build tools to see what fine-tuning changes inside them.

### 2026 — Tools for my own work

Using AI agents every day raised questions I wanted to answer by building: how to write down what an agent is allowed to do, how to stop it calling work done before tests pass, and how to carry decisions from one session to the next. Alongside that I still do storage and networking work, including a fix that was merged into an OpenWrt package.

# Timeline

## Foundations

### Grade 8 — Getting into security

I started poking at networks and security in grade 8 in Lahore, long before my first software job.

**Type:** Experience · **Workstream:** Security

**Technologies / skills:** Network security, Security testing.

Evidence: From my own notes.

### 2006 — First formal vulnerability assessment

In 2006, while still at school, I did my first formal vulnerability assessment, on a company's website, and found issues that would have let attackers in. The company's CEO later wrote me a recommendation letter.

**Type:** Experience · **Workstream:** Security

**Technologies / skills:** Vulnerability assessment, Security reporting, Client communication.

Evidence: Recommendation letter, October 2006.

## 2007–2012

### 2007–2011 — Computer-science education

Studied computer science at the Institute of Management Sciences in Lahore while working.

**Type:** Education · **Workstream:** Career

**Technologies / skills:** Computer science, Software engineering.

Evidence: From my own notes.

### May 2007–February 2009 — APTLOGIX: application development and integration

My first professional role included CMS and e-commerce development, C#/.NET backends and third-party integrations.

**Type:** Role · **Workstream:** Career

**Technologies / skills:** C#, .NET, PHP, E-commerce, SOAP/XML, API integration.

Evidence: Employment record.

### March–August 2009 — Senior Software Engineer at PureLogics

Worked as a senior software engineer before taking on a wider technical role.

**Type:** Role · **Workstream:** Career

**Technologies / skills:** Software engineering.

Evidence: Employment record.

### 2009–December 2012 — PureLogics: co-partner and CTO

Kept developing while also looking after network security and the company's systems.

**Type:** Role · **Workstream:** Career

**Technologies / skills:** Technical leadership, C, PHP, Python, Perl, Java, Systems administration, Network security.

Evidence: Employment record.

### From the PureLogics years — Virtualization, networks and telephony

Worked with VMware, Asterisk and Cisco networking. Later I moved from VMware to XenServer, and then to Proxmox/KVM.

**Type:** Experience · **Workstream:** Systems

**Technologies / skills:** VMware, XenServer, Asterisk, Cisco, Systems administration.

*Context: I don’t have exact start dates for these.*

Evidence: From my own notes.

### PureLogics period — A Facebook-integrated application

Worked on a community application that integrated with Facebook's APIs.

**Type:** Built · **Workstream:** Products

**Technologies / skills:** Application engineering, API integration, Product development.

Evidence: From my own notes.

### 2012–2013 — Co-founded Geeks Chicago

Co-founded a small consultancy and worked on client projects and solution design.

**Type:** Role · **Workstream:** Career

**Technologies / skills:** Entrepreneurship, Consulting, Solution architecture.

Evidence: From my own notes.

## 2013–2022

### January 2013–January 2022 — Xoho Tech: founder and CTO

Ran a consultancy that grew to about 40 people across Pakistan and the United States. I wrote code, designed systems, handled security and delivery, and did the sales.

**Type:** Role · **Workstream:** Career

**Technologies / skills:** Entrepreneurship, Software engineering, Solution architecture, Engineering leadership, Consulting, Delivery management.

Evidence: Employment record.

### Around 2013 — Starting with AWS

AWS became part of our client work early on.

**Type:** Experience · **Workstream:** Cloud

**Technologies / skills:** AWS, Cloud architecture.

Evidence: From my own notes.

### Xoho years — Digital archives and video platforms

Worked on AWS-hosted archive and video-streaming systems with searchable collections.

**Type:** Built · **Workstream:** Products

**Technologies / skills:** AWS, Video platforms, Archival systems, Web architecture, Search.

Evidence: From my own notes.

### Xoho years — Archive management

Built archive-management and audiovisual systems: data models, media handling and search.

**Type:** Built · **Workstream:** Products

**Technologies / skills:** Data modeling, Media integration, Archival systems, Search, Scalable architecture.

Evidence: From my own notes.

### Xoho years — Tools for archival quality and integrity

Worked on several applications for managing archive collections and checking data quality and file integrity.

**Type:** Built · **Workstream:** Products

**Technologies / skills:** Data quality, File integrity, Web applications, Product engineering.

Evidence: From my own notes.

### Xoho years — An e-signature application

Worked on an e-signature application on AWS: the signing workflows and integration with other business systems.

**Type:** Built · **Workstream:** Products

**Technologies / skills:** AWS, E-signature workflows, Web applications, Enterprise integration.

Evidence: From my own notes.

### Xoho years — Enterprise workflow automation

Built tools to automate quality-assurance workflows for a business.

**Type:** Built · **Workstream:** Products

**Technologies / skills:** Workflow automation, Enterprise integration, Software engineering.

Evidence: From my own notes.

### Xoho years — Writing down our engineering practices

Wrote the company's engineering guide: cloud practices, identity, secure coding, style and onboarding.

**Type:** Built · **Workstream:** Leadership

**Technologies / skills:** Engineering standards, IAM, MFA, Secure coding, Onboarding, Technical writing.

Evidence: From my own notes.

### April 2016 — AWS Solutions Architect Associate

My first AWS certification.

**Type:** Credential · **Workstream:** Career

**Technologies / skills:** AWS, Solution architecture.

Evidence: Certification record.

### Approximately 2016–2017 — Power cuts became a developer problem

Power cuts kept interrupting work and sometimes losing it. I started looking at hosting development environments remotely so work would survive an outage.

**Type:** Explored · **Workstream:** Developer tools

**Technologies / skills:** Developer experience, Operational resilience, Problem framing.

Evidence: From my own notes.

### Late 2010s — Taking the browser IDE apart

Looked at how VS Code separates the editor, the extension host and background services, then used browser-hosted VS Code in our later remote workspaces.

**Type:** Explored · **Workstream:** Developer tools

**Technologies / skills:** VS Code, Browser IDEs, Architecture, Developer tools.

Evidence: From my own notes.

### Late 2010s–2020 — A home lab to learn the layers underneath

Built a small Proxmox, Ceph and ZFS cluster at home to understand storage and virtualisation properly.

**Type:** Built · **Workstream:** Systems

**Technologies / skills:** Proxmox, Ceph, ZFS, KVM, High availability, Fibre networking.

Evidence: From my own notes.

### Late 2010s — Shared, template-based development environments

Built Linux-container templates with project dependencies and browser-based editing on a rented European server. Private networking, cloning and shared resources made the environment reusable across developers.

**Type:** Built · **Workstream:** Developer tools

**Technologies / skills:** Linux containers, Proxmox, VS Code, PHP, MySQL, Node.js, VPNs, Environment templates.

Evidence: From my own notes.

### Xoho years — Stopping idle workspaces

Used a web server and shell scripts to stop idle workspaces and start them again when a browser request came in, instead of keeping every environment running.

**Type:** Built · **Workstream:** Developer tools

**Technologies / skills:** Bash, Reverse proxies, Lifecycle automation, Resource optimization.

Evidence: From my own notes.

### Around 2019 — Firecracker developer environments

Moved the workspace experiments from containers toward Firecracker microVMs, weighing isolation against resource use and startup time.

**Type:** Experiment · **Workstream:** Systems

**Technologies / skills:** Firecracker, MicroVMs, Virtualization, Isolation design.

Evidence: From my own notes.

### Late 2019 — Teaching the team containers

Taught the team containers and some of the infrastructure under our applications.

**Type:** Teaching · **Workstream:** Leadership

**Technologies / skills:** Containers, Technical teaching, Engineering leadership.

Evidence: From my own notes.

### Around 2019–2020 — Workspaces for onboarding and handovers

Used saved editor state, snapshots, cloneable environments and preinstalled dependencies so people could hand over work, get started on a project, or come back to an old one.

**Type:** Built · **Workstream:** Developer tools

**Technologies / skills:** Snapshots, Developer onboarding, DevOps, Virtualization, Reproducibility.

Evidence: From my own notes.

### Around 2020 — Media-monitoring platform work

Worked on systems that collect, search and monitor media content.

**Type:** Built · **Workstream:** Products

**Technologies / skills:** Data ingestion, Search, Media platforms, Platform engineering.

Evidence: From my own notes.

### December 2017 — Planning an open platform for audiovisual metadata

I was on the core team planning an open platform that uses speech recognition, OCR and other machine learning to describe large audiovisual archives. In September 2017 we spent three days on requirements, comparing candidate tools for each part of the pipeline and sketching workflows. The plan was presented at the AMIA conference that December.

**Type:** Proposed · **Workstream:** Research

**Technologies / skills:** Audiovisual archives, Metadata, Speech recognition, OCR, Machine learning, Workflow design.

Evidence: Workshop notes, September 2017, and public project documentation.

[Project documentation](https://github.com/AudiovisualMetadataPlatform/amp_documentation/wiki)
[2018 report](https://hdl.handle.net/2022/21982)

## 2021–2022

### 4 June 2021 — Access to OpenAI's API beta

Got access to OpenAI's private API beta while running the consultancy, and started experimenting with it.

**Type:** Access · **Workstream:** AI & agents

**Technologies / skills:** Generative AI, LLM APIs.

*Context: This is the date access was granted.*

Evidence: From my own records.

### 2022 — Researching browser-based compute and developer tooling

Explored browser extensions, React internals, WebRTC desktop access, cloud isolation and customizable developer environments.

**Type:** Explored · **Workstream:** Developer tools

**Technologies / skills:** React, Browser extensions, WebRTC, AWS Nitro, Developer tools.

Evidence: From my own notes.

### February 2022–July 2024 — Cevo: AWS cloud consulting

Moved into cloud consulting in Australia, working across solution architecture, application integration and technical delivery.

**Type:** Role · **Workstream:** Career

**Technologies / skills:** AWS, Solution architecture, Cloud consulting, Technical leadership.

Evidence: Employment record.

### February 2022–July 2024 — Cloud API delivery and domain-team leadership

Built cloud APIs and integrations, including Consumer Data Right services, and led the engineers working on them.

**Type:** Built · **Workstream:** Cloud

**Technologies / skills:** AWS, API architecture, Domain-driven design, Data integration, Technical leadership, Observability.

Evidence: From my own notes.

### Cevo years — Enterprise search that respects permissions

Worked on enterprise search with Amazon Q Business and SharePoint, where results had to respect who is allowed to see each document.

**Type:** Built · **Workstream:** AI & agents

**Technologies / skills:** Amazon Q Business, SharePoint, Enterprise search, RAG, Access-aware retrieval.

Evidence: From my own notes.

### Cevo years — Document and email processing pipelines

Worked on pipelines that process documents and email, using dbt, machine learning and Amazon Bedrock.

**Type:** Built · **Workstream:** Cloud

**Technologies / skills:** dbt, Document processing, Machine learning, Amazon Bedrock, Data pipelines.

Evidence: From my own notes.

## 2023

### By June 2023 — WebRTC beyond the browser IDE

Tried streaming full desktops to the browser over WebRTC, for development and QA work that needed more than an editor and a shell. Looked at bandwidth, frame rate, audio and what the client device needed.

**Type:** Experiment · **Workstream:** Developer tools

**Technologies / skills:** WebRTC, Remote desktop delivery, Developer experience, Bandwidth analysis.

Evidence: From my own notes.

### June 2023 — Explaining LLMs and Transformers

Prepared presentation material on Transformers, attention, fine-tuning, prompts, agents and plugins.

**Type:** Teaching · **Workstream:** AI & agents

**Technologies / skills:** Transformers, Attention, Fine-tuning, Prompting, Agents, Technical teaching.

Evidence: From my own notes.

### 6 June 2023 — Writing about open language models

Wrote about open models like Falcon, which can be adapted and run outside a single vendor’s service.

**Type:** Writing · **Workstream:** AI & agents

**Technologies / skills:** Open LLMs, Model evaluation.

Evidence: From a public post.

[Public post](https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7071782270163955712)

### June 2023 — The development environment is part of DevOps

Wrote up the idea that development environments should be versioned and reproducible like any other build, with dependencies, prebuilds and tooling defined in the delivery process instead of a setup guide.

**Type:** Proposed · **Workstream:** Developer tools

**Technologies / skills:** Environment-as-code, CI/CD, Prebuilds, Dependency management, Developer onboarding.

Evidence: From my design notes.

### June 2023 — Separate the IDE from the application runtime

Sketched a workspace with separate containers for the editor and the application, possibly run on ECS or EKS.

**Type:** Proposed · **Workstream:** Developer tools

**Technologies / skills:** Container architecture, ECS, EKS, Separation of concerns.

Evidence: From my design notes.

### By June 2023 — Built workspace images with GitLab CI/CD

Adapted an existing workspace base image, added project dependencies in a Dockerfile and arranged automated environment builds on change.

**Type:** Built · **Workstream:** Developer tools

**Technologies / skills:** Dockerfiles, GitLab CI/CD, Image customization, Reusable components.

Evidence: From my own notes.

### June 2023 — Developer portals as a shared starting point

Proposed an internal developer portal for APIs, docs, examples and environment setup. Backstage was one option I looked at.

**Type:** Proposed · **Workstream:** Developer tools

**Technologies / skills:** Backstage, Internal developer platforms, API catalogs, Self-service provisioning.

Evidence: From my design notes.

### June 2023 — Extension licensing as a design constraint

Checked how custom IDE builds, extension marketplaces and their licence terms would affect the developer-platform design.

**Type:** Explored · **Workstream:** Developer tools

**Technologies / skills:** IDE extensibility, Software distribution, Licensing awareness.

Evidence: From my design notes.

### June 2023 — Keeping workspaces near shared services

Looked at keeping development close to shared services, controlling package access in one place, and keeping the workspace separate from the device a developer uses to reach it.

**Type:** Proposed · **Workstream:** Developer tools

**Technologies / skills:** Data locality, Network security, Dependency supply chain, Resource pooling, Remote development.

Evidence: From my design notes.

### June 2023 — Private AI coding assistance

Proposed connecting privately hosted models to the development environments, so code stayed private and people had less reason to use unapproved AI tools.

**Type:** Proposed · **Workstream:** AI & agents

**Technologies / skills:** Private LLMs, GPU inference, IDE integration, AI security, Developer enablement.

Evidence: From my design notes.

### June 2023 — AI preparing a Jira task before a developer starts

Sketched a workflow where a new Jira task starts a temporary development environment, and an AI prepares a plan and a first pass before the engineer takes over.

**Type:** Proposed · **Workstream:** AI & agents

**Technologies / skills:** Agent workflows, Task orchestration, Ephemeral workspaces, AI-assisted SDLC.

*Context: This was a proposal. It wasn't built as a whole.*

Evidence: From my design notes.

### June 2023 — Showing the plan and checking the result

In the same design, the engineer could see what the AI planned and what it actually did, check the work, and stayed responsible for the result.

**Type:** Proposed · **Workstream:** AI & agents

**Technologies / skills:** Human-in-the-loop design, Agent observability, Provenance, Workflow handoff.

Evidence: From my design notes.

### June 2023 — Runnable environments for pull-request review

Proposed opening each pull request in a running browser workspace, so a reviewer could try the change as well as read the code.

**Type:** Proposed · **Workstream:** Developer tools

**Technologies / skills:** Pull-request workflows, Reproducible environments, Quality gates, AI-assisted SDLC.

Evidence: From my design notes.

### June 2023 — Starting with a small pilot

Worked out what the developer platform would cost to run and to teach, and suggested a small pilot with runnable examples instead of building everything at once.

**Type:** Explored · **Workstream:** Leadership

**Technologies / skills:** Architecture trade-offs, Pilot scoping, Change management, Technical communication.

Evidence: From my design notes.

### 2023–2024 — A talk on our remote-development setup

Gave a talk on how our setup went from hosted workspaces to containers, prebuilt environments and AI-assisted workflows.

**Type:** Teaching · **Workstream:** Developer tools

**Technologies / skills:** Technical teaching, Gitpod, Dev containers, Kubernetes, DevOps, Architecture.

Evidence: From my own notes.

### 24 August 2023 — Small, domain-specific models on consumer hardware

Shared experiments fine-tuning small GPT-NeoX models for one task, such as drafting email, and running them on ordinary hardware.

**Type:** Experiment · **Workstream:** Research

**Technologies / skills:** GPT-NeoX, Fine-tuning, Local inference, Small models.

Evidence: From a public post.

[Public discussion](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7099086738823815168)

### 20 September 2023 — Article-to-podcast workflow in Python

Wrote a Python script that rewrites an article as a podcast script and reads it out with macOS text-to-speech.

**Type:** Built · **Workstream:** AI & agents

**Technologies / skills:** Python, Text-to-speech, Content transformation, Automation.

Evidence: From a public post.

[Public discussion](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7110023563620884481)

### 21 September 2023 — Browser-to-podcast automation with Shortcuts

Shared an iOS/macOS Shortcut that does the same from the browser, without a terminal.

**Type:** Built · **Workstream:** Products

**Technologies / skills:** Shortcuts, Workflow automation, Text-to-speech.

Evidence: From a public post.

[Public discussion](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7110023563620884481)

### 12 October 2023 — Thinking about small-model serving economics

Wrote about serving small models that sit idle most of the time, and what it costs to keep a model loaded in memory.

**Type:** Proposed · **Workstream:** Research

**Technologies / skills:** Inference economics, Resource scheduling, Small models.

Evidence: From a public post.

[Public post](https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7118123380234493952)

### 12–13 October 2023 — Local and edge AI as an architectural choice

Wrote about running smaller models closer to the data, for privacy and cost reasons.

**Type:** Writing · **Workstream:** AI & agents

**Technologies / skills:** Local inference, Edge AI, Privacy, AI architecture.

Evidence: From a public post.

[Public post](https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7118180502360006656)
[Public post](https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7118547811301814273)

### 24 October 2023 — A multi-agent AutoGen prototype

Built a GPT-4 and AutoGen notebook with roles such as account manager, architect, critic and engineer, with a person approving the plan before code ran.

**Type:** Prototype · **Workstream:** AI & agents

**Technologies / skills:** AutoGen, Multi-agent architecture, Python, Tool execution, Human-in-the-loop design.

Evidence: From a shared notebook.

### 24 October 2023 — Knowledge graphs, vector search and RAG together

Argued that graph structure and vector search work better together than vector search alone.

**Type:** Proposed · **Workstream:** AI & agents

**Technologies / skills:** Knowledge graphs, Vector search, RAG.

Evidence: From a public post.

[Public discussion](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7120802693845041152)

### 24 October 2023 — Prompt and SQL-injection risks in AI applications

Pointed out injection risks in an approach that let a model query data, especially where untrusted input reaches templates and tools.

**Type:** Writing · **Workstream:** Security

**Technologies / skills:** Prompt injection, SQL injection, AI security.

Evidence: From a public post.

[Public discussion](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7122335779665674240)

### 26 October 2023 — Private, multi-model agent orchestration

Wrote about AutoGen and Griptape as ways to combine several models in one agent system, including private or open models.

**Type:** Writing · **Workstream:** AI & agents

**Technologies / skills:** AutoGen, Griptape, Multi-agent orchestration, Private LLMs.

Evidence: From a public post.

[Public discussion](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7123165946977890304)

### 9 November 2023 — Using language models in honeypots

Suggested using language models to make honeypots more convincing and more useful for studying attackers.

**Type:** Proposed · **Workstream:** Security

**Technologies / skills:** Honeypots, Adaptive security, AI security.

Evidence: From a public post.

[Public discussion](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7125907723015073792)

### 12 December 2023 — Structured metadata for retrieval

Wrote about how RDF and structured metadata can help language models find the right information.

**Type:** Writing · **Workstream:** AI & agents

**Technologies / skills:** RDF, Metadata modeling, Retrieval.

Evidence: From a public post.

[Public post](https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7140275620294311938)

## 2024

### March 2024 — Public talk: a generative AI second brain on AWS serverless

Gave a talk at a Serverless Melbourne meetup on building a personal "second brain" on AWS serverless while keeping the cost low. Most of it does not need a large model or real-time answers, so documents are stored in S3, organised with small keyword and topic models, and searched with a vector database that sits on S3. Small quantized models handle the cheap steps, and retrieval with a larger model runs only for content I pick.

**Type:** Teaching · **Workstream:** AI & agents

**Technologies / skills:** AWS Lambda, Step Functions, Amazon S3, LanceDB, RAG, Quantized models, Knowledge management, Cost-aware architecture.

Evidence: Public recording and my slides.

[Talk recording](https://www.youtube.com/watch?v=pR-Z0Q0i4AI)

### 2024 — Local supervised fine-tuning

Fine-tuned Llama and OpenChat-class models locally with llama.cpp: preparing the training data, training, and checking the results.

**Type:** Experiment · **Workstream:** Research

**Technologies / skills:** llama.cpp, SFT, LoRA, Dataset engineering, Local inference.

Evidence: From my own notes.

### 19 March 2024 — Synthetic data, fine-tuning and quantization loops

Tried a loop where a model helps write and clean up training data for one task, then gets fine-tuned and quantized, and the cycle repeats.

**Type:** Experiment · **Workstream:** Research

**Technologies / skills:** Synthetic data, Fine-tuning, Quantization, Model evaluation.

Evidence: From a public post.

[Public discussion](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7175746712274145280)

### 21 March 2024 — Writing up AI-prepared developer workspaces

Wrote up the 2023 idea publicly: once requirements and Jira tasks exist, a model plans and prepares the work in an isolated environment, and the engineer still owns and reviews the code.

**Type:** Proposed · **Workstream:** AI & agents

**Technologies / skills:** Agent workflows, Jira integration, Human-in-the-loop design, Developer environments.

Evidence: From a public post.

[Public discussion](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7175808299538251776)

### 23 March 2024 — Personal assistants and audio knowledge workflows

Wrote about personal-assistant workflows built from open tools and automation that turn reading material into short written or audio briefings.

**Type:** Proposed · **Workstream:** AI & agents

**Technologies / skills:** Personal AI, Knowledge distillation, Audio workflows, Automation.

Evidence: From a public post.

[Public discussion](https://www.linkedin.com/feed/update/urn%3Ali%3AarticleSegment%3A%28urn%3Ali%3AlinkedInArticle%3A7169737324086476800%2C7169737325466238976%29)

### 28 March 2024 — Structured reasoning with smaller models

Tried the situation, task, action, result (STAR) structure to guide small quantized models through a task and a review of their own work.

**Type:** Experiment · **Workstream:** Research

**Technologies / skills:** Structured prompting, STAR reasoning, Quantized models, Agent workflows.

Evidence: From a public post.

[Public discussion](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7179159130325078016)

### 8 April 2024 — Training data with explicit reasoning and protection structure

Tried building fine-tuning data that spells out the reasoning steps, including steps meant to resist misuse.

**Type:** Experiment · **Workstream:** Research

**Technologies / skills:** NER, Training-data engineering, JSONL, Prompt workflows, AI security.

Evidence: From a public post.

[Public discussion](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7183002250003247104)
[Public post](https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7183002249244151809)

### 30 April 2024 — Models and serialized data

Asked how models behave when given serialized or binary data, and what security risks that opens.

**Type:** Writing · **Workstream:** Security

**Technologies / skills:** Serialization security, Model evaluation.

Evidence: From a public post.

[Public post](https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7191044016577126400)

### 27 May 2024 — Trying Cedar for policy-as-code

Wrote about experiments with Cedar and how hard it is to manage security policy at scale.

**Type:** Explored · **Workstream:** Security

**Technologies / skills:** Cedar, Policy-as-code, Security architecture.

Evidence: From a public post.

[Public discussion](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7200967618000797696)

### 3 June 2024 — Small models in serverless functions

Wrote about using ONNX to fit small-model inference into serverless functions such as Lambda.

**Type:** Proposed · **Workstream:** Research

**Technologies / skills:** ONNX, AWS Lambda, Small-model inference.

Evidence: From a public post.

[Public discussion](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7199160851122053121)

### 1 July 2024 — AI platform security is not application security

Wrote about the security gaps that stay in the application even when the AI platform underneath has good controls.

**Type:** Writing · **Workstream:** Security

**Technologies / skills:** Application security, GenAI security, Threat modeling.

Evidence: From a public post.

[Public discussion](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7213416918706192384)

### July 2024 — AWS Solutions Architect Professional

Passed the AWS Solutions Architect Professional exam.

**Type:** Credential · **Workstream:** Career

**Technologies / skills:** AWS, Solution architecture.

Evidence: Certification record.

### July 2024 — Snowflake SnowPro Core

Passed the SnowPro Core exam while working on data platforms.

**Type:** Credential · **Workstream:** Career

**Technologies / skills:** Snowflake, Data platforms.

Evidence: Certification record.

### August–October 2024 — Lyrebird Health: tech lead on a clinical AI product

Tech lead on a clinical AI product, covering the AWS setup, AI engineering and security.

**Type:** Role · **Workstream:** Career

**Technologies / skills:** Applied AI, AWS, Technical leadership, Clinical AI.

Evidence: Employment record.

[Public post](https://www.linkedin.com/feed/update/urn%3Ali%3AugcPost%3A7234048995386400768)

### 2024 — Clinical AI: evaluation and traceability

On the clinical product I worked on structured extraction, evaluation (including LLM-as-judge), prompt versioning, tracing each statement back to its source, and protecting sensitive data.

**Type:** Built · **Workstream:** AI & agents

**Technologies / skills:** Structured extraction, LLM evaluation, LLM-as-Judge, Prompt versioning, Provenance, Security architecture.

Evidence: From my own notes.

### 2024 — Early voice-AI evaluation

Took part in early testing of a voice AI product, which got me into realtime and streaming systems.

**Type:** Evaluation · **Workstream:** AI & agents

**Technologies / skills:** Voice-AI evaluation, Realtime systems, Test design.

Evidence: From my own notes.

### 2024 — Responsible-disclosure work in AI systems

Worked on AI-system security research and submitted a vulnerability report through a responsible-disclosure channel.

**Type:** Research · **Workstream:** Security

**Technologies / skills:** AI security, Responsible disclosure, Protocol analysis.

Evidence: From my own notes.

### 10 November 2024 — Prompt experiments with reasoning models

Shared what I noticed when testing reward-style wording in prompts for reasoning models.

**Type:** Experiment · **Workstream:** Research

**Technologies / skills:** Prompt experiments, Reasoning-model evaluation.

Evidence: From a public post.

[Public post](https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7261256868306563073)

## 2025

### Since March 2025 — Commonwealth Bank: Staff Platform Engineer

Joined Commonwealth Bank’s AI Acceleration squad.

**Type:** Role · **Workstream:** Career

**Technologies / skills:** Applied AI, Staff engineering, MLOps, Technical leadership, Production engineering.

Evidence: Employment record.

[Public post](https://www.linkedin.com/feed/update/urn%3Ali%3AugcPost%3A7305421036261449729)

### 2025–2026 — Taking AI applications to production

Work across development, data and infrastructure to take AI applications from early design into production.

**Type:** Built · **Workstream:** AI & agents

**Technologies / skills:** Applied AI, Data engineering, Production engineering, Architecture, Reliability.

Evidence: From my own notes.

### 2025–2026 — Identity and authorization for AI applications

Worked on identity and authorization between services, and between agents and their tools, using least privilege and avoiding shared secrets.

**Type:** Built · **Workstream:** Security

**Technologies / skills:** OAuth2, OIDC, Workload identity, Agent authorization, Security architecture.

Evidence: From my own notes.

### 14 April 2025 — Researching small-model reasoning with RL and SFT

Wrote publicly about using reinforcement learning and supervised fine-tuning to study reasoning in small models.

**Type:** Research · **Workstream:** Research

**Technologies / skills:** Reinforcement learning, SFT, Small models, Reasoning-model training.

Evidence: From a public post.

[Public post](https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7317351143611801600)

### 2025 — Security testing of AI applications

Tested AI applications for security problems, mainly where model output turns into application behaviour.

**Type:** Evaluation · **Workstream:** Security

**Technologies / skills:** AI red teaming, Application security, Security evaluation.

Evidence: From my own notes.

### June 2025 — Comparing audio-analysis approaches

Built a Gradio prototype to compare ways of analysing audio: telling speakers apart, timestamps, and interacting with what was said.

**Type:** Prototype · **Workstream:** AI & agents

**Technologies / skills:** Gradio, Audio AI, Speaker diarization, Multimodal evaluation, Rapid prototyping.

Evidence: From my own notes.

### June–October 2025 — Writing about capability and agent risk

Wrote about the security risks that come with more capable models and agents, and the trade-offs of using smaller, task-specific models.

**Type:** Writing · **Workstream:** Security

**Technologies / skills:** Agent security, Risk analysis, Model evaluation.

Evidence: From a public post.

[Public post](https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7341202015957544963)
[Public discussion](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7389645631243309056)

### 19 June 2025 — An AI hackathon team

Led a team at an internal AI hackathon, from the problem to a working demo.

**Type:** Recognition · **Workstream:** Leadership

**Technologies / skills:** Rapid prototyping, Applied AI, Team delivery, Technical leadership.

Evidence: From a public post.

[Public post](https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7341431711701901313)

### 2025–2026 — Mentoring, standards and engineering delivery

Besides hands-on engineering, I help with design discussions, mentor engineers and work on how the team delivers.

**Type:** Experience · **Workstream:** Leadership

**Technologies / skills:** Mentoring, Engineering standards, DevSecOps, Delivery management.

Evidence: From my own notes.

## 2026 · January–June

### 2026 — Security Champion

Took on the Security Champion role for my squad.

**Type:** Experience · **Workstream:** Leadership

**Technologies / skills:** Cybersecurity, Security advocacy, Risk awareness.

Evidence: From my own notes.

### 22–23 February 2026 — Declarative agents and filesystem-native tools

Suggested defining agents declaratively and exposing their tools as ordinary files and command-line programs.

**Type:** Proposed · **Workstream:** AI & agents

**Technologies / skills:** Agent architecture, MCP, OCI, Filesystem-native tools, Policy design.

Evidence: From a public post.

[Public post](https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7431312683196604416)
[Public discussion](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7423012527451561984)
[Public design discussion](https://github.com/agentskills/agentskills/discussions/179)

### February–March 2026 — Reasoning scaffolds for small models

Studied how small models learn when given reasoning scaffolds, and compared the insides of base and fine-tuned models.

**Type:** Research · **Workstream:** Research

**Technologies / skills:** GRPO, LoRA, MLX, Mechanistic interpretability, Experiment design.

Evidence: From a public post.

[Public post](https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7431243976348938240)

### March 2026 — Testing what model adaptation preserves and changes

Shared what I saw inside models after fine-tuning, and suggested checking whether it holds for other model families.

**Type:** Research · **Workstream:** Research

**Technologies / skills:** Activation analysis, Model evaluation, Cross-model evaluation, Fine-tuning.

Evidence: From a public post.

[Public discussion](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7433581529886412802)

### 12 April 2026 — Latent Contextual Reinforcement

Wrote up a research idea about guiding small models to reason better without retraining all of their weights.

**Type:** Research · **Workstream:** Research

**Technologies / skills:** Reasoning-model training, Expert scaffolding, Model evaluation.

Evidence: From a public post.

[Public post](https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7449012298184069120)

### By April 2026 — A Go proxy for models and tools

A single Go binary that offers an OpenAI-style chat API, routes to different model backends, and keeps tool registration and execution separate.

**Type:** Prototype · **Workstream:** AI & agents

**Technologies / skills:** Go, API compatibility, Tool execution, Agent loops.

Evidence: From my own notes.

### May–June 2026 — JuiceFS fork: storage, caching and backend engineering

Working in a JuiceFS fork on storage backends, caching, replication, integrity and block access.

**Type:** Development · **Workstream:** Systems

**Technologies / skills:** Go, Filesystems, Distributed storage, Caching, Replication, Encryption, iSCSI.

*Context: This is work in my fork, not an upstream JuiceFS release.*

Evidence: From a public post.

### 2026 — MLX model-inspection and comparison tools

Built command-line tools that show a model’s predictions layer by layer, compare activations between models, measure drift and run ablations.

**Type:** Development · **Workstream:** Research

**Technologies / skills:** MLX, Activation analysis, Logit lens, Model comparison, CLI/TUI, Mechanistic interpretability.

Evidence: From my own notes.

### June 2026 — A lesson in parallel-agent integration

Running several coding agents in parallel, I let them build on a shared interface before it was settled, and had to redo work. Now I agree on and test the shared interface first.

**Type:** Experience · **Workstream:** Developer tools

**Technologies / skills:** Multi-agent orchestration, Interface design, Dependency management, Integration testing.

Evidence: From my own notes.

### By June 2026 — CTXConfig: choose context at the symbol level

Tried choosing context for coding agents by symbol, using the syntax tree, instead of feeding in whole files.

**Type:** Prototype · **Workstream:** Developer tools

**Technologies / skills:** Rust, Tree-sitter, AST-aware context, Token budgeting, Agent tools.

Evidence: From my own notes.

### 12 June 2026 — agentic-agile: test-driven agent workflows

Published a Claude Code plugin where I approve the plan first, tests come before code, work goes through review, and a task is only done when its checks pass.

**Type:** Published · **Workstream:** AI & agents

**Technologies / skills:** TDD, Agent workflows, Claude Code hooks, Review gates, Developer tools.

Evidence: From a public post.

[Public post](https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7471139939909271552)

### 13 June 2026 — Measuring agent-assisted development

Reported the numbers from one agent-assisted development run (tokens, packages, tests, commits) instead of just saying it worked.

**Type:** Experiment · **Workstream:** Research

**Technologies / skills:** Agent evaluation, Token accounting, TDD, Engineering metrics.

Evidence: From a public post.

[Public post](https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7471380688865284096)

## 2026 · July–October

### 2026 — Running agents in microVMs

Ran packaged agents in microVMs with libkrun and compared isolation options against the same list of requirements.

**Type:** Prototype · **Workstream:** Systems

**Technologies / skills:** MicroVMs, libkrun, OCI, Agent isolation, Threat modeling, Virtualization.

Evidence: From my own notes.

### 3 July 2026 — AgentRC: writing down what an agent may do

Published a draft spec: a packaged agent lists the models, tools and network access it needs, and the platform can allow, narrow or refuse each one before it runs.

**Type:** Published · **Workstream:** AI & agents

**Technologies / skills:** AgentRC, OCI, Cedar, Go, BuildKit, Policy-as-code.

Evidence: From a public post.

[Public post](https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7478720397790687232)

### 5 July 2026 — Looking back at local-model work

Wrote a look back from early fastText and local-model work to later fine-tuning and defensive tool design.

**Type:** Writing · **Workstream:** Research

**Technologies / skills:** fastText, Local inference, Fine-tuning, Defensive tool design.

Evidence: From a public post.

[Public discussion](https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A7479004887133937664)

### 24–28 July 2026 — rclone filesystem adapters and media integration

Built filesystem adapters in an rclone fork, including one for a media server. The upstream proposal was closed without merging.

**Type:** Contribution · **Workstream:** Systems

**Technologies / skills:** Go, rclone, Filesystem adapters, Emby, Media integration, Open-source collaboration.

Evidence: From a public post.

[Upstream proposal](https://github.com/rclone/rclone/pull/9657)

### August 2026 — Versioned work state, not just conversation history

Designed a state document that records decisions, versions and task status across agent sessions, with an interactive view of how it changed over time.

**Type:** Proposed · **Workstream:** AI & agents

**Technologies / skills:** State modeling, Python, React, TypeScript, API architecture, Versioning.

Evidence: From my own notes.

### By August 2026 — Keeping my AI conversations

Built browser extensions to save my own AI conversations, and what they produced, as files I keep.

**Type:** Prototype · **Workstream:** Developer tools

**Technologies / skills:** Browser extensions, Data preservation, Conversation export.

Evidence: From my own notes.

### By August 2026 — Local Python execution inside a document editor

Prototyped a browser extension that runs Python locally with Pyodide, so code examples in a document can run where they are written.

**Type:** Prototype · **Workstream:** Developer tools

**Technologies / skills:** Python, Pyodide, WebAssembly, Browser extensions, Developer tools.

Evidence: From my own notes.

### 2026 — MLX reinforcement-learning and guided-training tooling

Developed training tooling for Apple Silicon, with configurable tasks, evaluators, rewards, supervised modes and guided reasoning experiments.

**Type:** Development · **Workstream:** Research

**Technologies / skills:** MLX, Python, GRPO, LoRA, SFT, Reward design, Training infrastructure.

Evidence: From a public post.

[Training repository](https://github.com/adeelahmad/mlx-grpo-trainer)

### 19–24 August 2026 — OpenWrt traffic-control fix merged upstream

Fixed bidirectional traffic shaping in luci-app-trafficctl, with tests. The pull request was merged upstream.

**Type:** Merged · **Workstream:** Systems

**Technologies / skills:** OpenWrt, Linux networking, nftables, tc/IFB, Traffic shaping, Regression testing.

Evidence: From a public post.

[Merged pull request](https://github.com/YusDyr/luci-app-trafficctl/pull/27)

### By August 2026 — Out-of-band automation through PiKVM

Prototyped a hardware-control interface for fixing and administering machines when the normal network is down.

**Type:** Prototype · **Workstream:** Systems

**Technologies / skills:** PiKVM, Hardware control, Recovery automation, Agent tools.

Evidence: From my own notes.

### September 2026 — Local GRPO and LoRA experiments on Apple Silicon

Kept training small models at home: guided-reasoning runs, quantisation and tooling to make runs repeatable.

**Type:** Experiment · **Workstream:** Research

**Technologies / skills:** MLX, GRPO, LoRA, Quantization, Dataset engineering, Experiment tracking.

Evidence: From my own notes.

### 2026 — ReqWall: request rules kept separate from enforcement

Tried keeping request-blocking rules in one place, so whatever enforces them, the browser or the router, can be swapped.

**Type:** Prototype · **Workstream:** Security

**Technologies / skills:** Policy modeling, HTTP analysis, Browser tools, Router integration, Security architecture.

Evidence: From my own notes.

### By September 2026 — A media-ingestion pipeline

Started a reusable pipeline that takes in audio, video and images and makes them searchable.

**Type:** Development · **Workstream:** Products

**Technologies / skills:** Media pipelines, Search indexing, Audio processing, Computer vision.

Evidence: From my own notes.

### September 2026 — Android device management

Checked what Android device management is actually possible, and made ADB and Shizuku start reliably.

**Type:** Experiment · **Workstream:** Systems

**Technologies / skills:** ADB, Shizuku, Android administration, Feasibility analysis.

Evidence: From my own notes.

### 16 September 2026 — agent-handoff: durable decisions across sessions

Published a Go tool that carries settled decisions, open tasks and versioned state from one AI session to the next, so I don’t rebuild context each time.

**Type:** Published · **Workstream:** AI & agents

**Technologies / skills:** Go, State versioning, Content-addressed storage, Agent handoff.

Evidence: From a public post.

[Public post](https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7505913245740744704)

### 19–22 September 2026 — Snapback: browse backup history without replacing live data

Designed a Linux/macOS tool to browse backup history read-only, folder by folder, and copy files back, without ever mounting over the live files.

**Type:** Proposed · **Workstream:** Systems

**Technologies / skills:** Go, FUSE, macFUSE, restic, rclone, Filesystems, Backup and restore.

Evidence: From my own notes.

### 23 September 2026 — aloop: the next version of agentic-agile

Reworked agentic-agile into aloop: test-driven agents, rounds of review, a final check, and a handoff file carried between fresh sessions.

**Type:** Development · **Workstream:** AI & agents

**Technologies / skills:** Agent engineering, TDD, Workflow design, Context handoff.

Evidence: From my own notes.

### 24–28 September 2026 — Workstation recovery and GPU diagnostics

Worked through memory compatibility, firmware settings, storage and network hardware, Linux drivers and GPU recovery on an older workstation.

**Type:** Built · **Workstream:** Systems

**Technologies / skills:** Hardware diagnostics, Linux, Proxmox, NVIDIA, Storage, Networking.

Evidence: From my own notes.

### 25 September 2026 — Browser storage and extension debugging

Exported data from IndexedDB and worked out why DOM and URL APIs behaved differently in page, extension and worker contexts.

**Type:** Development · **Workstream:** Developer tools

**Technologies / skills:** JavaScript, IndexedDB, Browser extensions, Data export.

Evidence: From my own notes.

### 27–30 September 2026 — Wireless backhaul and OpenWrt diagnostics

Investigated home wireless-backhaul constraints, routing and bridging behavior, TCP settings and radio state across a mixed-device network.

**Type:** Experiment · **Workstream:** Systems

**Technologies / skills:** OpenWrt, Wi-Fi, Routing, Bridging, TCP tuning, Network diagnostics.

Evidence: From my own notes.

### 29 September 2026 — A writing project on AI application principles

Began shaping a writing project about repeatable architecture, governance and operating principles for AI applications.

**Type:** Proposed · **Workstream:** Leadership

**Technologies / skills:** Technical writing, AI architecture, Governance.

Evidence: From my own notes.

### 30 September 2026 — USB-device emulation on Raspberry Pi

Started a USB device-emulation prototype on a Raspberry Pi Zero, possibly moving to other hardware later.

**Type:** Experiment · **Workstream:** Systems

**Technologies / skills:** USB OTG, Linux gadget interfaces, Raspberry Pi, Protocol analysis.

Evidence: From my own notes.

### 2 October 2026 — Privacy questions in voice-AI pipelines

Looked at how voice-AI pipelines handle personal information in streaming audio and voice tokens.

**Type:** Explored · **Workstream:** Security

**Technologies / skills:** Voice AI, Privacy, PII handling, AI security.

Evidence: From my own notes.

### 3 October 2026 — anytopdf: media to searchable documents

Designed a pipeline that turns files, images, audio and video into searchable documents, with transcripts, captions and keyframe notes. The first version is in Python; a later one may move to Rust.

**Type:** Proposed · **Workstream:** Products

**Technologies / skills:** Python, Rust, PDF processing, OCR, Video, Audio, Plugin architecture, RAG.

Evidence: From my own notes.

### 4 October 2026 — Lens: look and interface

Worked on the look and interface for Lens across web, app and command line.

**Type:** Development · **Workstream:** Products

**Technologies / skills:** Product design, Brand systems, React, Cross-platform interfaces.

Evidence: From my own notes.

### 4 October 2026 — Lens: a private AI hub

Planned a private AI hub that runs on my own machines: my notes and files, searchable, with answers that cite sources and actions that need my approval.

**Type:** Proposed · **Workstream:** Products

**Technologies / skills:** Local AI, RAG, Knowledge graphs, Hybrid search, Privacy, MCP, OAuth2, Automation.

Evidence: From my own notes.

### 5 October 2026 — Putting this record together

Put this timeline together to show how the work connects.

**Type:** Writing · **Workstream:** Leadership

**Technologies / skills:** Evidence modeling, Knowledge management, Data provenance.

Evidence: From my own notes.

## Cross-period

### Various dates — Security research at execution boundaries

Studied the sandbox boundaries of AI execution environments through protocol and systems analysis. Unpublished findings are not listed here.

**Type:** Research · **Workstream:** Security

**Technologies / skills:** Sandbox forensics, Protocol analysis, AI security, Execution-boundary analysis.

Evidence: From my own notes.

### Earlier years — Speech recognition, OCR and visual-media processing

Worked with speech-recognition, OCR and media-processing tools before the later LLM-focused projects.

**Type:** Experience · **Workstream:** AI & agents

**Technologies / skills:** Kaldi, DeepSpeech, Speech recognition, Tesseract, OCR, Computer vision, Media pipelines.

Evidence: From my own notes.

### Across the systems years — Telephony and realtime communication

Worked across Asterisk, FreePBX and later telephony systems, including audio codecs, recording, SMS integration and realtime browser communication.

**Type:** Experience · **Workstream:** Systems

**Technologies / skills:** Asterisk, FreePBX, 3CX, Opus, WebRTC, STUN/ICE, GSM/SMS.

Evidence: From my own notes.

### Across many projects — Routing, VPNs and identity-aware access

Worked with switching, routing, firewalls, overlay networks, VPNs and identity-aware access, picked up over many years and projects.

**Type:** Experience · **Workstream:** Systems

**Technologies / skills:** Cisco, MikroTik, OpenWrt, OSPF/BGP, VXLAN, Open vSwitch, WireGuard, IPsec, Zero Trust, Cloudflare.

Evidence: From my own notes.

### Undated — MacPilot: natural-language desktop automation

Experimented with model-driven macOS control using accessibility interfaces, AppleScript, OCR and asynchronous orchestration.

**Type:** Prototype · **Workstream:** AI & agents

**Technologies / skills:** Python, macOS Accessibility, AppleScript, OCR, Computer-use automation, Async orchestration.

Evidence: From my own notes.

## Projects

### AgentRC

Writing down what an agent needs, so a platform can decide what to allow.

Status: Published working draft.

### agent-handoff

Versioned decisions and work state across AI sessions.

Status: Published tool.

### agentic-agile / aloop

Test-first workflows, review stages and explicit handoff.

Status: Published plugin / evolving workflow.

### MLX training tooling

Configurable local reinforcement-learning and guided-training experiments.

Status: Development & research.

### Model inspection

Layer-wise predictions and comparison of adapted models.

Status: Research tooling.

### JuiceFS fork

Storage backends, caching, replication and block access.

Status: Fork development.

### OpenWrt trafficctl

Bidirectional shaping, diagnostics and regression testing.

Status: Merged contribution.

### rclone / Emby integration

Filesystem-backed media integrations.

Status: Prototype / unmerged upstream proposal.

### CTXConfig

Symbol-aware context selection for coding agents.

Status: Prototype.

### Snapback

Read-only backup history and copy-back restoration.

Status: Design.

### Lens

My own notes and files, searchable, with answers that cite sources and actions that need my approval.

Status: In development.

### anytopdf

Turning media and files into searchable documents.

Status: In development.

### MacPilot

Natural-language automation through macOS interfaces.

Status: Experiment.

### ReqWall

Portable request policies with replaceable enforcers.

Status: Prototype.

### Developer workspaces

Browser IDEs, environment images, microVMs and starting and stopping workspaces on demand.

Status: Experiments.

### Versioned agent state

Records of decisions and task state, not just transcripts.

Status: Design.

### PiKVM agent interface

Out-of-band recovery and hardware-control experiments.

Status: Prototype.

### Media ingestion

Reusable audio, video and image-processing pipelines.

Status: Development.

### Python in documents

Local browser execution with Pyodide and WebAssembly.

Status: Prototype.

## Everything I’ve used

Skill tags show what a piece of work involved. They are not ratings.

### Programming & software

C, C++, C#, .NET, Go, Rust, Python, PHP, Java, JavaScript, TypeScript, Ruby, Perl, Bash, Lua, Swift, AppleScript, React, Angular, Node.js, Deno, Spring, Symfony, Laravel, Rails, SQLAlchemy, Browser extensions, Design patterns, Testing, Metaprogramming.

### AI, agents & model research

Transformers, Attention, RAG, Knowledge graphs, FAISS, Qdrant, pgvector, LangChain, AutoGen, ReAct, MCP, Amazon Bedrock, Amazon Q Business, MLX, LoRA, GRPO, RLHF, SFT, LLM evaluation, LLM-as-Judge, Prompt versioning, Local inference, Embeddings, Mechanistic interpretability, Speech recognition, Speaker diarization, Kaldi, DeepSpeech, Tesseract, OCR, Computer vision.

### Data, search & caching

MySQL, PostgreSQL, MariaDB, MongoDB, IndexedDB, LevelDB, MQTT, Apache Solr, OpenRefine, InfluxDB, Airflow, Metabase, Redash, Redis, Memcached, FlashCache, Varnish, Squid.

### Storage & filesystems

Block storage, Object storage, FUSE, Overlay filesystems, NFS, iSCSI, Samba, ext filesystems, HFS, APFS, XFS, ZFS, GlusterFS, Ceph, Btrfs, JuiceFS, NBD, SCSI, Snapshots, Replication, Cloning, Storage caches.

### Cloud & delivery

AWS, Azure, Google Cloud, OpenStack, OpenNebula, CircleCI, GitHub Actions, GitLab CI/CD, CodeBuild, CodePipeline, Travis CI, KMS, SSM Parameter Store, Docker BuildKit, OCI, Engineering metrics, Technical leadership, Developer onboarding.

### Virtualization & containers

QEMU, KVM, Proxmox, VMware, XenServer, libvirt, Firecracker, Cloud Hypervisor, rust-vmm, LXC, LXD, Docker, Docker Swarm, Podman, Kata Containers, Kubernetes, GPU passthrough, PCI/USB passthrough, High availability, AppArmor, cgroups, Container plugins.

### Networks, Linux & identity

DNS, mDNS, DNS-over-TLS/HTTPS, Linux networking, Linux kernel modules, Ansible, HAProxy, Apache, NGINX, Caddy, Open vSwitch, VXLAN, FRRouting, OSPF/BGP, L2/L3 switching, Fibre networking, WireGuard, OpenVPN, IPsec, Tailscale, ZeroTier, Nebula, Cloudflare, Zero Trust, LDAP, Active Directory, OpenLDAP, SSO, Kerberos, RADIUS, SAML, Keycloak, SSH tunnels, iPXE, Netboot, DHCP, TCP tuning.

### Routers & firewalls

Cisco, MikroTik, OPNsense, pfSense, UniFi, OpenWrt, iptables, nftables, Vyatta, Traffic shaping, VPNs.

### Observability & operations

Prometheus, Grafana, ELK, Logstash, Nagios, OpenNMS, Cacti, SmokePing, mtr, iperf, mtail, Monit, NetFlow, SNMP, Fail2ban, Tripwire, Lynis, Log analysis.

### Voice & device systems

Asterisk, FreePBX, RasPBX, 3CX, Opus, WebRTC, STUN/ICE, GSM/SMS, Kannel, Call recording, Amazon Connect, Apple MDM, PiKVM, ADB, Shizuku, USB OTG, Raspberry Pi.

### Security & forensics

Threat modeling, Vulnerability assessment, Security reporting, Secure coding, OSINT, Network packet analysis, Data recovery, Browser forensics, Protocol analysis, TLS inspection, Frida, Binary analysis, Firewall hardening, Mitmproxy, OpenSSL, Certificate authorities, AI security, Prompt injection, DLP, OAuth2, OIDC, RBAC, Audit logging, Sandbox forensics, Responsible disclosure.

## About this page

This is a selection, not everything. Work for employers and clients is described in general terms, without client names or internal details.

Some entries link to public posts or code. The rest come from my own notes and records.

Some dates are exact and some are approximate. Approximate ones say so.

Skill tags show what a piece of work involved. They are not ratings.

Experiments and proposals are labelled as such. Not everything here was finished or went into production.