Obsidian
Built a real-time inference pipeline for a Series B fintech — handles 12k req/s of fraud scoring with sub-50ms p99 latency. Replaced a brittle rules engine with a fine-tuned model behind a gRPC gateway.
Selected Work
Recent projects where we helped teams ship AI-native products. Each one involved real constraints, real users, and tight timelines.
Structured Retrieval Engine
Designed a hierarchical RAG system for a legal-tech company. 400k documents indexed with custom chunking that preserves clause structure. Cut hallucination rate from 18% to under 3%.
Cross-lingual QA Pipeline
Built a multilingual question-answering system across 14 languages for an enterprise SaaS platform. Accuracy holds within 4% of English-only baselines even on low-resource languages.
The Quiet Terminal
Redesigned a developer-facing AI tool from chatbot UI to structured workspace. Task completion time dropped 40% once we replaced free-form prompting with constrained, schema-driven inputs.