Services
I build custom AI agents for startups and SMBs, with human approval, tests and live monitoring from day one. Live in 3–6 weeks, at a fixed price.
Both are fixed price after a free scope week. Not sure which one? That is what the scoping call decides.
Deep-dive on your workflow, constraints, and success metric. I write a one-page scope with explicit non-goals. You sign off before any billing starts.
Framework choice (LangGraph vs CrewAI vs custom), state schema, tool list, eval strategy, cost projection. You approve before a line of code is written.
Weekly demo of a working slice. You see tool calls, traces, and eval scores every Friday — not a black box.
Docs, training, deploy runbook. Then I'm on call for bug fixes and tuning for a full month after launch.
My default stack is LangGraph for anything stateful, Claude for reasoning-heavy agents, OpenAI for high-throughput tool-calling, and LangSmith for tracing from day one. I always ship an eval harness before I ship the agent — un-evaluated agents are how you find out about regressions from your customers instead of your CI.
Real questions from scoping calls. If yours isn't here, ask it on a free 30-minute call.
$15–30K for a Vibe-Built MVP that proves the agent with real users in about 3 weeks, or $30–60K for an AI Product Sprint: a production agent with human-in-the-loop approval, an eval harness and tracing, in 4–6 weeks. Both are fixed price, agreed after a free scope week, so there are no hourly surprises.
A validated MVP ships in about 3 weeks. A production system with human-in-the-loop, an eval harness, and multi-agent orchestration takes 4–6 weeks. Either way, the free scope lock comes first, so you know the timeline before any billing starts.
A workflow follows a fixed path — step A, then step B, then step C. An agent decides what to do next based on what it just observed. Agents are powerful when the path isn't known up front (messy customer queries, open-ended research tasks); workflows are cheaper and more predictable when it is. I usually recommend a workflow unless the work genuinely needs dynamic decision-making.
Three layers: structured tool outputs over free-text where possible, an eval harness that scores every change against a fixed test set, and human-in-the-loop checkpoints on high-risk actions.
Ask three things. Who writes the code: the person on the sales call, or a team you never meet? Is the price fixed before work starts, or billed by the hour? And do you own the code, prompts and evals at the end? Many top AI agent development companies are strong on the first call and weak on the handoff. With me, one architect scopes, builds and hands over everything in your GitHub.
Default to a single agent. Multi-agent systems are useful when you have clearly separable specialties (a researcher, a writer, a critic) or latency benefits from parallelism, but they introduce coordination bugs and extra tokens. I only recommend multi-agent when the supervisor-worker decomposition is obvious.