The Engineering Notebook — page 2 of 38
Notes on building with AI, agents & the modern stack.
Deep dives on AI/ML, RAG systems, agent engineering, and senior-engineer architecture decisions.
Technology REA: Reverse Engineer Anything With AI Agents [2026]
REA is the open-source MCP server that lets Claude Code, Codex, and Cursor reverse-engineer binaries, Electron apps, APKs, and live browsers. Here's what it does, how to set it up, and the legal line nobody else is drawing.
AI and Machine Learning How to Evaluate EmbeddingGemma 2 for Multilingual RAG [2026]
EmbeddingGemma 2 is shiny. Shipping it into a multilingual RAG stack without regressions is not. Here’s an embeddings-only eval + rollout playbook that works.
Developer Tools OpenAI Decisions API Tutorial [2026]: 3 Cost-Safe Patterns
A production-first OpenAI Decisions API tutorial: typed evals on POST /v1/decisions, predicate/choice/score, and three patterns to gate, batch, and escalate without cost blowups.
Developer Tools How to Build a Local Model Registry for GGUF [2026]
Stop downloading random weights from links. Build a local model registry with GGUF versioning, provenance attestations, eval gates, and a real rollback plan.
Cybersecurity How to Stop Prompt Injection in Agent Web Search APIs [2026]
Web search tools turn every AI agent into a browser. Here’s a threat model and a safe reference design: allowlists, extraction, sandboxing, signed provenance, and citation rules.
Cybersecurity How to Detect AI-Generated Bug Bounty Reports [2026 Guide]
A maintainer-first playbook to score report quality, spot LLM-written spam, add proof-of-work friction, rate-limit intake on GitHub, and reproduce safely in a sandbox.
AI and Machine Learning How to Run Qwen 3.8 Flash Next 125B in Strata on RTX 4090 [2026]
A reproducible Strata setup for Qwen 3.8 Flash Next (125B) on RTX 4090, plus the KV cache knobs that decide whether you hit ~100 tok/s or hit OOM.
AI and Machine Learning How to Tell If You Need an AI Agent or a Workflow (7 Smells) [2026]
Most “AI agents” in production are deterministic workflows wearing an LLM costume. Here’s the teardown rubric I use to simplify designs, cut cost, and make on-call sane.
AI and Machine Learning Kolibri Open Weight Model Benchmark: 1M Context Reality Check [2026]
Aleph Alpha’s Kolibri ships downloadable weights under Apache-2.0 and claims a 1M-token context. Here’s what that changes for EU/on-prem teams and how to benchmark it reproducibly on an A10/A100.
Cybersecurity How to Build a Prompt Injection Scanner for RAG [2026]
A practical, reproducible way to scan retrieved RAG context for indirect prompt injection, benchmark false positives, and gate regressions in GitHub Actions without blocking your team all day.
Developer Tools ds4 dwarfstar [2026 Review]: Redis-Style Local LLM Loops
ds4 (DwarfStar 4) turns local LLM inference into a Redis-like workflow: one long-lived server, deterministic prompt-keyed KV caching, and fast edit-run-edit loops without a GUI circus.
Developer Tools How to Follow arXiv Updated Rate Limit Policy [2026]
arXiv tightened rate limits on Oct 1, 2026. Here’s a compliant ingestion blueprint for RAG: incremental harvest, caching, backoff, bulk S3, WARC snapshots, and version tracking.