AI Engineer — agents, context engineering, RAG with a human in the loop

I design the system around the model, not the prompt. The decisions that matter are which documents enter the context window, when an extra agent is worth its cost, what the model is deliberately not allowed to see, and where a person confirms before a decision counts.

Computer Engineering, Instituto Federal Fluminense (2026). Almost two years at Petrobras, in logistics, putting an LLM inside a real decision process.

• Context scoping — what enters the window, what stays out, and why leakage between cases invalidates the output and the experiment measuring it
• Agent topology — isolated context so parallel agents don't anchor on each other; conditional routing so the expensive agent only runs when it earns its place
• Human confirmation — the system proposes with reasoning and sources; a person approves before the result is authoritative, and every decision stays traceable to the criterion behind it
• Honest evaluation — measuring whether retrieval actually helped, and reporting the cases where it changed nothing
• AWS serverless when the thing has to actually run — Lambda, S3, DynamoDB, Cognito, SQS, CDK

Multi-agent grading of discursive exam answers, modelled on the ENEM process of two independent graders plus a referee.

Two examiners score the same answer in isolated context — neither sees the other's score or reasoning — and a third agent is invoked only when their divergence crosses a threshold. RAG is hard-scoped per exam with no global fallback, because context leaking between exams would contaminate both the grading and the study measuring it.

Controlled study, 24 graded answers. RAG improved evaluative specificity on mid-range answers (Δ = +0.84) and changed nothing at the extremes. The referee never fired: real divergence peaked at 1.22 against a threshold of 2.0, which says the threshold needed empirical calibration rather than a guess. Weak answers failed the stability criterion. Both negative results are in the README, along with what I would do differently.

LangGraph · LangChain · FastAPI · ChromaDB · PostgreSQL

Media ingestion and review platform for a video production studio, replacing ad-hoc file sending with a traceable flow. Native iPhone app, web review portal, serverless backend. Deployed to AWS and validated on a real device.

The design constraint that shaped everything: upload and publication are independent operations — putting a file in the cloud never exposes it to the client. Multipart uploads go straight from the phone to S3 through pre-signed URLs in resumable 16 MiB parts, so the file never travels through the API. Derivation runs async behind SQS with a dead-letter queue and idempotent Lambdas, so a retry can't duplicate an upload.

42 backend tests, 35 native iOS tests, infrastructure validated by CDK synth.

Swift · TypeScript · CDK · Lambda ARM64 · DynamoDB · S3 · Cognito (OAuth 2.0 + PKCE) · SQS · MediaConvert · CloudFront

Monthly audit of stock-discrepancy justifications submitted by a partner company. The coordinator read each justification and accepted or rewrote it by hand; closing the indicator took hours.

I started with a syntactic adherence indicator between the submitted text and the coordinator's corrected version. The RAG stage came later, once it was clear that the history of accepts and rejects was the context that was missing: the system retrieves precedents from 100+ validated cases and meeting minutes, proposes a 0–10 score with its reasoning, and an analyst confirms before the decision enters the corpus.

Time spent on that indicator fell from ~3 hours/week on average to at most 2–10 minutes.

It only became part of the routine once the team trusted the output enough to bring it into a live negotiation meeting. That took sitting with the analysts to find where their bar was stricter than the code's — not tuning the prompt.

Coursework and experiments from earlier on, kept because they're part of the record: PCA implementation, a neural network that drives a car. They are what they look like — things I built while learning.

Open to AI engineering / agent engineering roles, in Python. Remote, or hybrid and on-site in Rio de Janeiro.

Studying for the AWS Solutions Architect Associate, exam booked for September 2026.

LinkedIn