Work

A cobbled Victorian warehouse street in London, with iron overhead walkways linking the brick buildings and a single figure walking between them.

The index of what I build, write, and publish. Long-form essays live here on the site; technical write-ups are on dev.to, where they're canonical, and are collected below.

Open-source build: RoastPilot, a coffee-roasting agent

An agentic control system that roasts coffee on a real Hottop roaster, built to explore how much autonomy an agent should have when a wrong decision can affect a physical process that is difficult to reverse. The design principle throughout is that the deterministic controller owns the roast loop, while the large language model (LLM) only advises. A hard safety policy returns typed verdicts (allow, clamp, reject, emergency stop) before anything reaches a 200°C machine, and advisor prompts are pinned by replay bake-offs against logged real roasts, scoring drop-decision F1, heat / fan error, and latency. It has roasted every batch of my coffee since November 2025.

The RoastPilot device console: roast curve with event markers, heat and fan controls, and the LLM advisory panel with typed recommendations gated by allow verdicts
End of a roast on the device console: first crack detected through the Model Context Protocol (MCP) server, a drop recommendation at 0.95 confidence, and every advisory gated by an allow verdict.
  • roastpilot-agent: the deterministic agent harness and web console; a state-machine controller owns every hardware command, and the LLM advisor returns typed recommendations only.
  • coffee-roaster-mcp: spec-driven MCP server exposing fourteen tools for roast session lifecycle, telemetry, first-crack detection, and log export behind a single hardware boundary. On PyPI and the MCP Registry.
  • coffee-first-crack-detection: fine-tuned Audio Spectrogram Transformer detecting first crack from audio (98% accuracy, 0.93 macro F1), quantised to INT8 ONNX for torch-free inference on a Raspberry Pi 5. The model, dataset, and a live demo Space are on Hugging Face, with current metrics and benchmarks maintained on the cards.
  • roastpilot-plan: the cross-repository planning hub holding the agent orchestration architecture, decision records, and epic tracking.

The rebuild is spec-driven: repository rules, externalised plan state, and replay evals, with coding agents doing the implementation across more than 460 merged pull requests while I keep ownership of architecture, ML decisions, and quality gates.

The RoastPilot rig mid-roast: the Hottop roaster, the first-crack microphone, and a live agent session driving it over MCP
Roast day: an agent session drives the roaster over MCP while the first-crack rig listens.

Essays

Long-form essays on what agentic AI changes in software engineering, grounded in the systems I build and the evidence they produce. New essays appear here automatically as they are published.

Verification Has a Clock · 31 Aug 2026
The cost and timing of being wrong play a key role when deciding on agent autonomy

The Factory Was the Real Invention · 19 Jul 2026
The factory, not the machine, was the real invention. The final post in the series: how the organisation decides what AI productivity is worth, and who it is for.

Better Ground, Not a Better Model · 1 Jul 2026
You don't get a better agent by buying a better model alone. You get one by preparing the ground it runs on: the tests, the structure, the rules it runs inside.

Cheap Tokens, Cheap Attempts · 7 Jun 2026
Cheap generation doesn't make engineering judgment obsolete. It makes it the scarce thing, and the first Industrial Revolution shows why.

Read the three-part AI and the Industrial Revolution series in order.

Writing on dev.to

Technical posts are published on dev.to/syamaner and remain canonical there. The most recent:

Spec-driven ML: rebuilding the coffee agent (2026)

A six-part series on rebuilding the prototype through a multi-phase machine-learning project, directing coding agents while keeping ownership of architecture, model decisions, and quality gates. It ends with a live roast: four agents, one MCP server, and a supervised run on real hardware.

  1. The architecture and the agent
  2. Building the audio dataset
  3. The science: tuning to high precision
  4. Optimising an 86M-parameter audio transformer for Raspberry Pi
  5. From local model to live demo on Hugging Face
  6. Roast day: four agents and a live roast over MCP

The original prototype (2025)

How the project started: training the first-crack detector, building MCP servers for roaster control, and closing the loop with an agent.

  1. Training a neural network to detect first crack from audio
  2. Building MCP servers to control a home roaster
  3. Orchestrating MCP servers with .NET Aspire and n8n

Applied AI and RAG with .NET Aspire (2024–2025)

Earlier engineering writing (2022–2024)

Research and teaching

In 2024, I completed an MSc in Artificial Intelligence at the University of Bath with distinction. Before moving into industry, I conducted doctoral research in augmented reality at the Pattern Recognition and Image Analysis group, University of Salford, and chose an engineering career rather than completing the PhD.

Sharing what I learn has remained part of my engineering work. I've been a Kubernetes trainer at ASOS, run TDD and machine-learning kata sessions at Flagstone that twice won a culture award, delivered technical-transformation talks across European markets representing Mercedes-Benz Cars UK, and run hands-on agentic development sessions for engineering teams.

  • Data Mining Approach to Implement a Recommendation System for Electronic Tour Guides. EEE 2005.
  • Mining GPS Logs to Augment Location Models. WIT Transactions on Information and Communication Technologies, 2005.
  • Determining the Locations Visited by GPS Users: A Clustering Approach. CISST '04.

Profile and outputs: Semantic Scholar · Salford repository

Elsewhere