Stéphane Hamel

Platform engineer. I build agentic SDLC pipelines and the infrastructure they run on.

Thesis — bounded autonomy. Let agents do the work, but enforce in code what a prompt can only ask, ground every verdict in executed evidence, and keep a human at the irreversible step. Every repo below is that idea under a different constraint.

Agentic SDLC

autonomous-dev-loop	Issue → AI PR → evidence-backed AI review → auto-fix → human merge, on GitHub Actions. Built after an agent replaced a 26-test suite with an 18-line stub; 11 of its own features were shipped by the loop. 800+ tests, 25 ADRs.
ai-dev-tools	The interactive counterpart in Claude Code: typed builders, an adversarial challenger, and a reviewer that runs your tests before it calls a patch done.
agent-patterns	15 multi-agent patterns indexed by the problem they solve — diagrams, trade-offs, failure modes, runnable code. 12 extracted from the two repos above.

Infrastructure

gitops-homelab	Chat, code, vision, speech-to-text and text-to-speech models on one 6 GB GTX 1060 — bare-metal microk8s, ArgoCD, GPU time-slicing. 25 ADRs, 13 incidents written up with root cause.
ollama-chat	Self-hosted chat UI for local LLMs, deployed by the repo above.

Platform fundamentals

sync-engine	Webhooks drop events, polling is late: a Python sync engine built around 11 integration failure modes, each mapped to an ADR and a test.
fullstack-pilot	Node, Python and .NET on MongoDB, PostgreSQL and SQL Server — with CI that tests database image upgrades against existing data.

flowchart LR
    subgraph sdlc["Agentic SDLC"]
        adl["autonomous-dev-loop<br/>headless · GitHub Actions"]
        adt["ai-dev-tools<br/>interactive · Claude Code"]
        ap["agent-patterns"]
    end
    subgraph infra["Self-hosted infrastructure"]
        gh["gitops-homelab<br/>microk8s · ArgoCD · GTX 1060"]
        oc["ollama-chat"]
    end
    subgraph fund["Platform fundamentals"]
        se["sync-engine"]
        fp["fullstack-pilot"]
    end
    adl -- "patterns extracted to" --> ap
    adt -- "patterns extracted to" --> ap
    adl -.-|"same pipeline, two modes"| adt
    gh -- "deploys" --> oc

Loading

How I work

• Decisions in writing. 65+ ADRs across these repos, each with rejected alternatives and consequences.
• Incidents, not hypotheticals. Guardrails and config trace back to a specific failure: a deleted test suite, a CUDA OOM, a Helm setting that never reached the pod.
• Measure, then decide. Quantization chosen by benchmark and by executing the generated code, not by reputation.
• Honest scope. What's enforced in code and what's only asked of a model are documented separately.

Current focus

• Moving LLM guardrails from prompts into deterministic checks — import allowlists and type-check gates before a PR opens (ADR-0019, proposed).
• Evaluating local models on constrained hardware by running what they produce (ADR-0011).

Node.js · .NET · Python · Kubernetes · ArgoCD · GitHub Actions · Claude API · Ollama