//Resume//Resume//Resume

Built a trace-generation pipeline producing post-training data for a 31B-parameter open-weight LLM, reviewed SFT, distillation, and LoRA literature, and supported a senior engineer’s Google Cloud training runs. Contributed to an internal RecursiveMAS reimplementation as a product prototype. Reworked a LangGraph/LangSmith harness, cutting token usage 50% and improving accuracy by about 5% on selected models. Built a 77-task, human-verified evaluation suite and prototyped Qwen3.5-4B inference on iPhone with MLX, streaming responses to AR glasses.

Built a Python computer vision pipeline to process C-SPAN videos and score emotions and reactions to congressional speeches. Structured LexisNexis coverage by outlet, syndication, quote placement, and timing, pairing LLM-generated tone labels with committee-led human review. Combined video reaction scores and article text for multimodal analysis of ethos, pathos, and logos.

Created an interactive 3D space environment with Three.js and React Three Fiber, synchronizing rocket motion and altitude-based transitions with page scrolling. Built and deployed an authenticated platform for a 50-quest mentorship program with Turso/libSQL persistence, role-authorized uploads, transactional emails, and PostHog analytics and error monitoring.

Supported conference technology for 55,000 attendees at UNFCCC COP29.

Applied exploratory data analysis and feature engineering to taxi-driver data, correlating ride and cash-out patterns with income and building monthly-income prediction models. Worked with boosting models and prototyped a Python credit-scoring pipeline using bank records and alternative signals. Built an internal LLM assistant over Payriff API documentation for developer Q&A.