About Caio Theodoro
Senior AI Engineer with 7+ years in software engineering and 3 years building production LLM systems — the agent infrastructure Caio Theodoro builds, and the roles behind it.
I am Caio Theodoro, a Senior AI Engineer: 7+ years in software engineering and 3 years building production LLM systems. My work is the infrastructure agents run on — execution harnesses, human review, evaluation and observability, policy guardrails, and the multi-model inference layer underneath. Python and TypeScript, deep on RAG, structured outputs, agentic orchestration, and the token-cost engineering that decides whether an AI feature is economically viable.
What I work on
At Adopt AI I architected the core agent execution harness — turn and message lifecycle, idempotent claim semantics, replay-safe Temporal orchestration — which every capability the platform ships runs on: human-in-the-loop review, policies, evals, generative UI. I built the live evaluation dashboard that scores conversation quality in production so degradation surfaces before customers feel it, and the HITL intervention system end to end: a signal-driven resolve state machine with optimistic locking and tenant isolation, wired into live agent workflows over a Temporal signal contract.
The generative UI architecture for tax and accounting workflows is the piece I am most pleased with. Instead of letting the model emit markup each turn, the harness selects from a registry of approved components — that cut dashboard prompts from roughly 15k tokens to 700–2k, and a JSON-Patch delta protocol saved 90% on multi-turn follow-up edits. Alongside it, a policy engine for guardrails with version pinning and audit logging, and an AWS Bedrock integration that put multi-model inference behind one selection interface.
Before that
I founded Avenza in February 2025 and ran it as an AI consulting and product studio alongside full-time work: agents, computer vision systems, analytics platforms and automation delivered to outside clients. I directed a 7-engineer team across discovery, solution architecture, delivery and engineering standards, on systems tied to roughly $2.8M USD in client revenue over eight months. I advise the studio now rather than run it day to day.
Before Adopt AI I was a Senior Software Engineer at Acorns, working through a phased Rails-monolith-to-services migration at a consumer fintech serving millions of users — Node services, React, a GraphQL gateway, and the ephemeral per-PR environments that closed a staging/production parity gap. Earlier, at MB Labs, cloud-native fintech, education and supply-chain systems on AWS with event-driven pipelines under compliance constraints.
I write down what I measured. Every project on this site comes with the numbers that justified it, including the ones that did not go my way.
Selected projects
- ReconForge — A laptop-trained 1.7B LoRA for financial reconciliation exceptions: 0.913 severity-weighted recall against a frontier model’s 0.872, at zero API cost.
- Suture — An 8B vision-language model that diffs an underwriting binder against the issued policy — 0.959 recall across 13 discrepancy classes where zero-shot frontier vision scores 0.373.
- LossBench — Scores money-touching agents on expected operational loss instead of accuracy: severity-weighted, calibrated, replayable, on a hash-chained decision ledger.