Now

One big thing this month: a self-learning proxy debugger at the day job. HAR diff engine plus an LLM + RAG pipeline that proposes config patches for proxy failures, validates them through the same CI/CD path a human fix would take, and — if the fix holds — writes it back into a vector database so the next occurrence of the same failure class gets resolved without touching the LLM at all. Staleness tracking, human gate on complex fixes via Telegram, blue-green rollback protection. Fixes that compound.

In parallel: Argus Intelligence's first product — the EU AI Footprint Scanner — is live and selling via Lemon Squeezy. Trial signups and support are quiet-day work; the debugger is where the build time goes.

Shipped

• EU AI Footprint Scanner — Argus's first product, now live and selling via Lemon Squeezy. The GitHub App runs on every pull request and posts findings as a Check Run plus a summary comment; the CLI still ships alongside it.

Building

• Self-learning proxy debugger (day job) — HAR diff plus LLM + RAG for automated config-patch proposals and validation. Design is settled; this month is closing the loop so verified fixes become the retrieval corpus for the next failure. Full writeup: I built a CI/CD pipeline that fixes itself — and remembers how.
• Local LLM benchmarking — off the bench: quantized Qwen2.5-72B-Instruct to Q6_K on the Mac Studio and published the numbers — the compression cost nothing measurable (+0.0057 PPL, within noise; decode 2× faster). Next: the Q4_K_M run and KL-divergence. The 4B/8B TPS sweep from the 16GB machine is still drafted, not published.

Planning

• PRD locked for an Argus Compliance Copilot — RAG over the EU AI Act that explains scanner findings against the regulation, planned as a paid tier. Closed beta targeted for early 2027.

On the bench

Personal-project work that hasn't materially advanced since Argus launched. Named honestly rather than hidden — I'll pick these back up once the debugger stabilises.

• aegis-rag — RAGAS eval harness is still the missing piece.
• YOLO custom detection — label audit paused; the vision-model work at the day job on the QA gate is filling the same learning slot.
• Rotation service — Phase 1 + Phase 2 running in production; HMAC + proactive probe + Prometheus metrics deferred behind the debugger.

Learning

• Vector-DB knowledge staleness — how to invalidate cached knowledge against a codebase that keeps changing, without burning an LLM call every lookup. Commit-hash-anchored fix metadata is the current answer.
• Reranker architectures — when a cross-encoder is worth the latency tax vs. just retrieving more candidates.

Not doing

• Training foundation models. Fine-tuning giant models on rented GPUs isn't where my time has the best return right now — the retrieval and inference-serving side is.

Inspired by Derek Sivers' /now page convention.