2024 — Present
Belo Horizonte, BR (Remote)
Founder & AI/ML Engineer · Magoquiz
- Built a self-hosted LLM scoring service that runs about 200x cheaper per call than the hosted API. Distilled a Sonnet judge into a fine-tuned Llama 3.1 8B (LoRA/QLoRA on Modal), compiled it to a TensorRT-LLM engine behind NVIDIA Triton on a scale-to-zero L4 GPU, serving at ~$0.0003 per score (P50 1.3s).
- Made the judge provider-portable across AWS Bedrock (Converse API, IAM, cross-region inference profiles) and a self-hosted GPU behind one interface, switchable with a single environment variable.
- MLOps/LLMOps: a CI quality gate (GitHub Actions) that blocks AI quiz-generator regressions with an LLM-as-judge on every relevant PR, a Weights & Biases model registry, and a daily drift monitor over Langfuse traces that opens a ticket when quality drops.
- Ran a pre-registered experiment over 13 model/prompt setups against an 88-quiz benchmark tied to 130k sessions. Scored each with an LLM-as-judge calibrated to real conversion, used paired significance tests, and confirmed the production setup was already cost-optimal. Published a public W&B report.
- Built a hybrid-search RAG (dense + BM25 fused with Reciprocal Rank Fusion) in LlamaIndex, benchmarked on hit-rate@5 and MRR. Hybrid beat dense-only by 20pp.
- Drove 47x ROI and a 236% conversion lift in 10 days for a strategic client, through an OAuth 2.0 (PKCE) + webhooks integration with Yampi, Resend, and PostHog.
- Built the full-stack platform solo on a four-piece architecture (edge-deployed SvelteKit on Cloudflare Workers, a background-job worker, a token-refresh service, and a workflow runner) on PostgreSQL, serving 30k requests a day for 18 months with no customer-facing outages.
- Shipped 9 e-commerce integrations on reusable OAuth + webhook abstractions (about 70% less time per integration), a programmatic-SEO pipeline of 2,500+ AI-reviewed posts (+300% organic impressions), and 230+ tests (unit, integration, E2E) behind a CI/CD pipeline.