AI engineer who ships. I build agentic systems, RAG pipelines and the evaluation harnesses that prove they actually work — not notebooks that die in a folder.

AI Engineer Intern — Innovun Global · Aug 2026 – Present

• Develop a production RAG assistant serving Cruz Roja Mexicana (Red Cross, Mexico) across 3 channels.
• Debug live failures by tracing production logs and telemetry to the component at fault before shipping a fix.
• Work with cross-functional teams and client stakeholders across time zones, authoring the technical documentation that keeps delivery repeatable.

AI Engineer Intern — AI LifeBOT · Jan – Jun 2026

• Built 3 production LLM applications and autonomous agents using LangChain + OpenAI, serving 200+ users.
• Architected end-to-end GenAI pipelines with RAG, vector databases, and real-time streaming, achieving 35% lower latency and 99.5% uptime.
• Developed multi-agent orchestration with approval gates, telemetry, and audit logging.

Machine Learning Research Intern — CFM, RCOEM · May – Jul 2025

• Built a clinical prediction system using 1,000+ patient records with biometric feature engineering, achieving 87% accuracy.
• Engineered measurable features by hand from raw, inconsistent tabular data in Python and pandas.
• Reduced development time by 30% through pipeline standardization, and raised result reliability 25% via cross-validation across multiple splits.

Catches silent production failures — stale upstreams, creeping null rates, renamed columns — before a KPI moves. Runs a scan → score → incident → write-back → remediate loop over DataHub via MCP, naming the exact downstream models at risk and generating fail-fast remediation code. Ships with bundled fixtures: runs with no Docker and no API keys, one env var away from live.

Python FastAPI MCP DataHub Gemini GitHub Actions

A deterministic simulation environment for benchmarking autonomous email agents across classification, prioritization and full action execution, with bounded, numerically stable grading metrics. Four policy modes (heuristic, stress-test, LLM-driven, multi-agent hybrid), episode replay, approval workflows — and published benchmark results with honest findings.

Multi-Agent RAG FastAPI React SciPy PostgreSQL

Two-stage pipeline — two-tower embeddings → ANN retrieval → LightGBM ranker — with a full MLOps loop: event bus, feature updater, online store, model registry, retrain jobs. Interface-first design puts every dependency behind a Protocol, so it runs CPU-only and Windows-native with zero external services, then scales to Kafka/K8s by flipping one config value.

PyTorch LightGBM MLflow Kafka Redis Prometheus

20+ more — agentic platforms, India-market AI, Web3 and IoT — on my portfolio and repos.

Open to AI/ML engineering roles. Reach me at rogerdemello289@gmail.com