Practical, cited write-ups on AI agents, evaluation, security and cost — the trade-offs we hit on real client work. 14 articles and counting.
1 article
2 articles
/ Document AI
When to read invoices, forms and contracts with multimodal LLMs, when classic OCR still wins, and how to validate extraction before it hits your ERP.
13 min readRead
/ Security
A step-by-step plan for testing tool-using AI agents against direct and indirect prompt injection, data exfiltration and over-broad permissions.
13 min readRead
2 articles
/ Engineering
How to roll out Claude Code, Codex, Cursor or Copilot across a team: permissions, sandboxes, instruction files, MCP, CI guardrails and measuring impact.
11 min readRead
/ Integration
What it takes to ship a Model Context Protocol server customers can trust: task-shaped tools, OAuth-based authorization, rate limits, evals and versioning.
13 min readRead
1 article
1 article
2 articles
/ AI AgentsUpdated October 2026
Budgets, retries, idempotent tools, approvals, prompt-injection defences, durable execution, OTel tracing and evals: how we keep AI agents reliable in 2026.
15 min readRead
/ Fine-TuningUpdated October 2026
When to prompt, retrieve or fine-tune in 2026: decision ladder, caching costs, SFT vs DPO vs RFT, who offers what, and a worked cost comparison.
11 min readRead
2 articles
/ SecurityUpdated October 2026
What Bedrock, Azure, Model Armor and OpenAI guardrails actually check in 2026, where each one stops, and the agent controls you still have to build.
16 min readRead
/ ArchitectureUpdated October 2026
When multiple agents beat one, five patterns that survive production, the single-writer rule, MCP vs A2A, and an even-handed look at 2026 agent frameworks.
13 min readRead
2 articles
/ InfrastructureUpdated October 2026
pgvector, OpenSearch, Pinecone, Qdrant, Weaviate, S3 Vectors and more: hybrid search, filtering, quantization, evals and costs at 1M–100M vectors.
13 min readRead
/ Cost OptimizationUpdated October 2026
How to cut and predict Amazon Bedrock spend in 2026: cost attribution, prompt caching, model routing, batch and service tiers, with a worked cost example.
13 min readRead
