The EU AI Act deadline moved: the engineering did not
What the EU AI Act asks engineers to build — risk classification, Article 50 transparency, data lineage and inference logs — and why a moved deadline is a bad reason to start.
$ git log --oneline posts/
Field notes.
Long-ish posts about work we shipped: the decisions, the bugs, the numbers. No think-pieces, no listicles.
Most recent
Migrating to the stateless MCP spec: what the 2026-07-28 revision actually costs
What the MCP 2026-07-28 revision requires, how to move per-session state out of a production server, and an honest account of what the migration costs.
KV cache is a cache: configure its failure modes
Prefix caching is shipped as a throughput switch, but it is a cache with a hit rate, an eviction policy and a cold path that quietly governs your p99 latency.
The embedding ceiling and the reranker budget
Why upgrading your embedding model stops paying, what a cross-encoder recovers that a single vector cannot, and how to size a reranker candidate budget on your own data.
Inference savings are a procurement decision, not a kernel problem
The arithmetic behind self-hosting versus an API, what low-precision serving actually buys, and which inference optimisations survive the next vendor price cut.
The problem with enterprise NL-to-SQL is specification, not model capability
Why text-to-SQL still returns confidently wrong numbers in production, what a governed semantic layer has to contain to stop it, and how to evaluate the result honestly.
Agent memory: graph versus vector was the wrong argument
What LoCoMo actually scores, why the published memory numbers moved by 25 points, and why write policy and invalidation are the parts no benchmark grades.
Speculative decoding at production concurrency: a latency trick, not a throughput one
Why speculative decoding speedups measured at batch size one shrink or invert under continuous batching, what acceptance rate actually controls, and how to decide.
The agent harness: build the loop, adopt the sandbox
The agent harness is four build-versus-adopt decisions wearing one name. Where owning the loop pays for itself, and where it quietly does not.
Durable execution for agents: replay or re-decide
Durable execution engines assume that re-running your code reproduces the same decisions. LLM agents break that assumption, and the recovery semantics you pick have real costs.
Chance-corrected agreement: your LLM judge is worse than your dashboard says
Raw agreement gives an LLM judge credit for guessing. Why Cohen's κ belongs on the dashboard instead, what it does not fix and how to run the position swap.
Your retry loop voided your conformal guarantee
Conformal prediction gives a real coverage guarantee, but it assumes exchangeability and one shot per item. Retries, drift and calibration reuse quietly break it.
Reading order is the document AI metric that matters
Character-level accuracy hides the failure that actually breaks document RAG. Why reading order deserves to be your headline parsing metric, and how to measure it without labels.
Streaming video models: the hard part is knowing when not to respond
Streaming video-language models are graded on answer accuracy, but live deployments fail on response timing, false alarms and frame budget. What to measure instead.
Iceberg v3 row lineage: the equality-delete trap waiting in your CDC consumer
Iceberg v3 makes row lineage mandatory, but the spec exempts rows updated via equality deletes — and that gap quietly corrupts any consumer treating _row_id as a stable key.
MCP servers are a supply-chain decision, not a dependency install
Third-party MCP servers put attacker-controlled text into your model context. What the disclosed incidents actually show, and the dull controls that work.
Choose a vector database on renewal risk, not recall
Why ANN recall stopped being the differentiator, and how to evaluate a vector store on exit cost, filter portability and re-embedding cost instead — with the test to run first.
Read everything, write through a gate: permissions for operational agents
Why we give operational agents broad autonomous read access and force every write through a pull request, what that costs during an incident and where we carve exceptions.
Document AI: when a sub-1B specialist beats a frontier model
Cost per page, p99 latency, deployability and fine-tunability decide document pipelines — and on those metrics a 0.9B specialist usually wins, until your document distribution has a tail.
Browser agents: the honest ceiling on unattended web automation
Published web-agent benchmarks put state-changing task success below 50 percent. The compounding arithmetic behind that ceiling, and how we design around it.
MCP in Production: Building Agentic AI Systems That Actually Ship
A practical guide to deploying Model Context Protocol agents in production — from tool registration and multi-agent orchestration to observability and the mistakes we made along the way.
Streaming-First Data Architecture: Why We Replaced Batch ETL with CDC + Iceberg
How we migrated a Fortune 500 company from 24-hour batch ETL to a streaming-first lakehouse with Debezium CDC, Apache Flink, and Apache Iceberg — and what broke along the way.
How We Built a Production Hallucination Detection Pipeline
An end-to-end walkthrough of the multi-agent LangGraph system we built to verify 30K+ content pieces daily at 92% factuality accuracy.
Graph Neural Networks for Real-Time Fraud Detection at Scale
Why rule-based fraud systems miss coordinated attacks, how we built a GNN-powered detection engine processing 2.3M transactions daily at sub-800ms latency, and the tradeoffs we navigated.
Building an MCP-Powered Restaurant Intelligence Platform from Scratch
How we unified messy restaurant data, built MCP agent tools for NL-to-SQL and demand forecasting, and delivered a 56% forecast accuracy improvement.
The Agent Stack in 2026: From Chatbots to Autonomous Systems
How the AI industry moved from chatbot demos to production agent systems — and what the new stack looks like with MCP, A2A, multi-agent orchestration, and FinOps.
From Four Tools to One Dashboard: Engineering a Solar Proposal Platform
How we consolidated a fragmented sales workflow into a unified proposal dashboard that cut creation time by 62% for a leading solar roofing company.