Field notes on building AI that ships
Costs, framework trade-offs, and hard-won lessons from putting AI agents, voice AI, and RAG into production. Written for the people who have to make the build-or-buy call.
What a custom AI agent costs in 2026 — from $1,500–$5,000 for single-task bots up to $150,000+ for multi-agent systems, plus run costs and ROI by scope.
A custom AI voice agent costs $5,000–$25,000 to build plus ~$0.05–$0.35/min to run in 2026 — build tiers, per-minute economics, and ROI.
A production architecture for AI agents with LangGraph: state graphs, the plan–act–verify loop, tool calling, human-in-the-loop, guardrails, and evals.
Most RAG chatbots fail on retrieval, not the LLM. The 8 fixes — hybrid search, reranking, query rewriting, citation verification, and evals — that work.
Vapi, Retell, and Bland are voice-AI orchestration platforms; Twilio is the telephony beneath them. How they compare on latency, control, and cost.
LangGraph vs CrewAI vs AutoGen for production AI agents in 2026: LangGraph for reliability, CrewAI for fast prototypes, AutoGen now in maintenance mode.
Pinecone vs Qdrant vs pgvector for RAG in 2026: pgvector if you're on Postgres, Qdrant for open hybrid search, Pinecone for zero-ops managed.
Why AI agents break in production — no guardrails, no evals, no observability, runaway loops, and bad tool design — plus a practical fix for each.
An engineer’s end-to-end guide to a real-time voice AI agent: telephony, STT, turn-taking, the LLM, TTS, barge-in, human handoff, and the latency budget.
What AI workflow automation costs in 2026 — per-workflow build ranges for n8n, Make, Zapier, and custom Python, plus the factors that move the price.
Let's see if I can take this off your plate
Tell me what you want to automate. On a free 30-minute call I’ll tell you straight whether it’s worth building, roughly what it costs, and how I’d approach it — no pitch, no obligation.
Book my free 30-min AI scoping call →