Field notes on
building AI in production.
The permanent library — long-form engineering and architecture writing on enterprise AI, MCP & agentic operations, platform engineering, and cloud modernization. LinkedIn is where the discussion happens; this is where the thinking lives.
RAG or fine-tuning? One agent, four configurations, one catalog change.
Qwen3.5-4B on an M5 Max: base, base + RAG, LoRA fine-tuned, both — then the org changed.
Fine-tuning alone scored 100% on the first snapshot. Then checkout moved teams; the adapter kept routing to the old channel 4/4 while both retrieval configurations followed the edit 4/4. Start with base + RAG, build a golden set, add fine-tuning when it earns its place. Deterministic scoring, every result in the repo.
Read the experiment →Zero Token Architecture, seen from MCP
Infer once. Export the logic. Run without inference — at runtime, where the tool is the export, and at build time, where the SDK is.
Kelsey Hightower's ZTA read as a design rule instead of a cost argument. Why most MCP servers are API wrappers with a nicer schema, what “curated” concretely means for a tool, where the four token costs of a call actually go, and the quieter build-time version: three servers that each re-derived the same auth from scratch, and the SDK export that runs the loop once. Plus a four-stage discovery→export→run→change lifecycle and a four-question test for which side of the line a tool sits on.
Read the article →A2A Joins AAIF: The Agentic Stack Settles — and Identity Is Next
Five layers, five jobs, one neutral home — and one real retail–finance order run through every layer to prove the composition.
Agent2Agent is now an AAIF project, next to MCP, goose, AGENTS.md, and agentgateway. Why the 2025 "A2A vs MCP" debate was a category error, what each layer actually does, and a concrete end-to-end walkthrough — signed agent cards, RFQs over A2A through the gateway, budget cleared by a peer agent, one human approval, payment via a scoped AP2 mandate, and an audit chain by construction. Plus the editorial: identity & trust is the missing sixth layer, and the working group to watch.
Read the analysis →Qwen3.8-27B dropped, and the AI community claims it beats the OG — Opus 4.6
A 27B on a laptop vs February's frontier — 24 real on-call tasks, deterministic scoring, $1.61.
The agentic claim, independently tested: identical pass rates on both tiers, the same trap failed by both models both runs, two different tool-calling personalities exposed by counting calls instead of passes — and a harness bug that nearly scored the whole thing wrong. Adversarial tier authored by a third model to kill authorship bias; every transcript reproducible.
Read the eval →MCP Just Went Stateless — the Biggest Redesign Since Launch
The 2026-07-28 revision removes the handshake, kills sessions, and stops servers from ever calling clients back — the same tool call went from nine messages to four.
A production engineer's walkthrough of MCP's first true architectural rewrite: no initialize, no Mcp-Session-Id, no server-initiated requests. Mid-request input now rides the Multi Round-Trip Request pattern — an HMAC-sealed requestState the client mails back — so any replica can finish the job and sticky sessions disappear. Sampling and roots deprecated. Plus the governance angle: the first revision shipped since MCP moved under the Agentic AI Foundation. Animated old-vs-new comparison included.
Inside the Forward Pass — What Actually Happens When an LLM Answers You
The forward pass taken apart stage by stage on a real quantized model — five animated stages, every number verified against the library's own kernels.
Tokenize → embed → attend → vote → loop, each stage re-done with raw matrix ops on Qwen3-0.6B under MLX and checked bit-for-bit: the recomputed attention matched the library kernel with max diff 0.0, the full 28-block pass reproduced model(x) exactly, and the KV cache ran 3× faster than the naive loop — plus the honest footnote where bf16 rounding flipped a greedy argmax. Runs on any Apple Silicon Mac with a 350 MB model.
Read Internals #1 →Agents Fail at the Seams — Field Notes from MCP Dev Summit Mumbai 2026
Two days, 16 sessions, one through-line: multi-agent AI fails between agents, tools, and protocols — and every failure is a missing contract.
A practitioner's synthesis of the summit: the RAM/disk/syscall mental model, the six spines, and a failures→contracts table where every fix is discipline, not a smarter model — security as the floor (confused-deputy authorization), durable context as the ceiling (the empty “event clock”), the tool surface as the product (TDQS), and the unowned MCP↔A2A seam. Plus the gaps worth building and verified ecosystem notes.
Read the field notes →MCP Access to a Database — as the User, Not a Service Account
Principal propagation for AI agents — carry the user's identity to the data source, not a shared service account.
A working demo: an AI agent reaches a Postgres banking database through MCP as the end user — role-based PII masking, per-tool RBAC, and audit attributed to the real human — plus a Postgres-18 path where the database engine itself denies unauthorized columns via SASL OAUTHBEARER. Honest throughout about what's app-layer versus engine-enforced.
Read the case study →How Far Can One Mac Go? Local LLM Limits on an M5 Max
Six models, 3B → 123B, on 128GB of unified memory — measured, charted, reproducible.
A live dashboard of the full ladder through one MLX benchmark harness on an M5 Max: the scaling curve, the memory ceiling that wasn't where Apple said (123B ran in 69GB; the probe hit 128GB with no refusal), and the 14-inch thermal tax — decode holds, prefill pays. Every number reproducible from the open-source repo.
Open the dashboard →SAP NorthStar AI Architecture: Context Moat, HANA's Elevation, and the Gatekeeping Problem
A practitioner's page-by-page read — what it gets right, and which doors stay gatekeeped.
Why the context thesis is the most honest thing an ERP vendor has said about AI, why HANA's move from in-memory database to context engine matters, and where the open-by-design rhetoric collides with data toll roads, ECC reality, and CI/CD.
Read article →More articles in this category are on the way.
More in progress — MCP framework patterns, AIOps agent design, and cloud-modernization field notes. Follow on LinkedIn to catch them first.