Now

Argus Intelligence's first product — the EU AI Footprint Scanner — is now live and selling. With that shipped, my build time this month has swung back to the systems side: the rotation service that health-checks and load-balances the proxy fleet.

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

• Rotation service — current focus is HMAC signature validation on the /node-health endpoint. It's designed but not yet enforced, so any caller with network reach can currently push health updates. That's the next thing I'm shipping.
• aegis-rag — wiring up a RAGAS eval harness so I can stop saying "it retrieves well" and start posting faithfulness numbers.
• Local LLM benchmarking — full sweep of Qwen and Llama variants at Q4_K_M / Q5_K_M / Q6_K on 16GB. The goal is a small honest table, not a leaderboard.
• YOLO custom detection — label audit pass before any more training runs. Dataset work, not hyperparameter work.

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.

Learning

• Reranker architectures — specifically when a cross-encoder reranker is worth the latency tax vs. just retrieving more candidates.
• Quantization math: why Q4_K_M outperforms plain Q4_0 per-byte, and what "K-quant" is actually doing.

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.