AI Development
Building with LLMs: agents, RAG, evals, and shipping to prod. 20 posts.
Surviving LLM API Failures: Timeouts, Fallbacks, Queues
LLM APIs fail for hours each quarter. Timeout budgets, retry rules that honour Retry-After, provider fallbacks, and queues that keep your app standing.
AI DevelopmentA Realistic AI Pair-Programming Workflow
A realistic AI pair-programming workflow — decompose to verifiable tasks, verify like a reviewer, and the three signals to take the keyboard back.
AI DevelopmentOpen-Weight Models in 2026: What's Actually Usable
Which open-weight models are production-usable in 2026, task by task, plus licence gotchas and the maths for when they beat APIs on cost and privacy.
AI DevelopmentChoosing a Vector Database: pgvector Until It Hurts
Start with pgvector and move only when one of four numbers turns red. The decision ladder to Qdrant, plus the benchmarks that matter and those that lie.
AI DevelopmentShipping AI Features Behind Flags
Ship AI features behind four controls — kill switches with defined fallbacks, shadow mode, cost circuit breakers, and pinned versions you can roll back.
AI DevelopmentSynthetic Data That Doesn't Collapse Your Model
How to generate synthetic training data without model collapse — diversity seeding, MinHash dedup, judge calibration, and the real-data anchor ratio.
AI DevelopmentPrompt Injection: Threat Model and Practical Mitigations
Prompt injection can't be prompted away. The threat model, the defences that fail under red-teaming, and the capability confinement that limits damage.
AI DevelopmentMultimodal Pipelines: OCR, Vision, and Audio in One Stack
How to build one pipeline for OCR, vision, and audio — routing per modality, engine choice with real costs, and fusing outputs into one JSON shape.
AI DevelopmentAI Code Review: Where It Helps, Where It Lies
Which bug classes AI code review catches, which it misses, and the false-positive budget that decides whether your team keeps reading its comments.
AI DevelopmentBuilding an MCP Server: Field Notes
Field notes from building MCP servers — tool design agents can use, output token discipline, stdio vs HTTP transports, auth, and multi-client testing.
AI DevelopmentLocal-First AI Apps: Patterns and Pitfalls
Where local-first AI works in 2026 — model sizes per task, runtime capability probes, sync rules for derived data, and when to escalate to the cloud.
AI DevelopmentTracing and Observability for LLM Apps
Logs can't debug a six-step agent. Traces with spans for prompts and tools, cost per request, sampling strategies, and self-hosting Langfuse.
AI DevelopmentContext Window Management: Budgeting, Caching, Compaction
Big context windows moved the problem from fit to cost. Set token budgets per feature, lay prompts out cache-first, and prune before you summarise.
AI DevelopmentEmbeddings Search: From Cosine Demo to Production
From cosine demo to production search: structure-aware chunking, hybrid BM25 plus vectors, cross-encoder reranking, and the eval set that proves it.
AI DevelopmentStructured Output From LLMs: Schemas, Grammars, Retries
Reliable JSON from LLMs is a solved problem: native structured outputs, grammar-constrained decoding for local models, and validation-retry loops.
AI DevelopmentAgent Architectures in 2026: Loops, Graphs, and Handoffs
Single loop, subagents, handoffs, or graphs — how to pick an agent architecture in 2026, and when multi-agent complexity actually pays for itself.
AI DevelopmentEvals Before Vibes: Testing LLM Apps Properly
Skip the dashboards. Build a 50-example golden set, write code assertions first, calibrate an LLM judge against human labels, and gate CI on it.
AI DevelopmentPrompt Engineering Is Just Engineering
Prompts are config that ships to production. Version them in git, test them in CI, review the diffs, and stop the recurring "model got worse" mysteries.
AI DevelopmentTool Calling Patterns That Don't Fall Apart in Production
Tool calling patterns that survive production: flat schemas, idempotency keys, parallel call handling, and error results the model can repair.
AI DevelopmentRAG vs Fine-Tuning: When Each Actually Wins
RAG injects knowledge, fine-tuning changes behaviour. A decision framework based on data volatility, latency, and cost, plus hybrid patterns that work.