Work
Production GenAI delivery across startups, banking, and open-source software
Strong in retrieval, evaluation, orchestration, and applied NLP systems
Remote-first collaboration across teams in the U.S., Europe, and Latin America
Multi-agent sales assistant for a 30M+ customer bank
Cut customer acquisition cost by 6x while improving sales team productivity.
The team needed a production AI assistant that could guide sales flows reliably across tools, stakeholders, and business constraints.
- Led the end-to-end build of a generative AI sales assistant using LangChain and LangGraph.
- Applied tool use, personas, workflows, guardrails, and model selection strategies to improve reliability.
RAG optimization for lower cost and better answers
Reduced inference cost by 90% while improving answer quality beyond competing LLM alternatives.
The existing RAG system was too expensive and needed stronger answer quality for a production banking context.
- Researched new GCP capabilities and prompt strategies.
- Reworked retrieval and prompting for stronger accuracy and cost efficiency.
Product retrieval redesign for startup search
Cut indexing time from hours to minutes and reached 90%+ recall with only 20 candidates instead of 100.
The product search engine needed materially better retrieval efficiency and fresher indexing over proprietary catalog data.
- Analyzed product groupings to redesign indexing strategy.
- Improved vector retrieval with Gemini-based data augmentation.
Flagship proof
Napolab
A benchmark for Portuguese language models that challenged assumptions about the advantage of Portuguese-only models and highlighted contamination and evaluation quality issues.
Hashformers
Transformer and beam-search based hashtag segmentation system recognized as state of the art in LREC 2022.
Selected writing
The Hidden Truth About LLM Performance: Why Your Benchmark Results Might Be Misleading
There are two systematic issues with benchmarks, which were already present in the community before the advent of LLMs but became worse …
Start with work, projects, or the resume.
Work, projects, research, and the resume cover the strongest proof first.