🛠️ Projects

Ran an 80-experiment autoresearch-style sweep across Notion's AI search embedding pipeline (millions of pages, 10M+ workspaces), tuning Ray Data + self-hosted embedding throughput.

3.2x wall-time speedup, 60% GPU cost reduction, and stable 10M-row processing - unlocked the ~250 → ~700 credits/day expansion across all Notion workspaces.

Built an agent-at-scale system over 7 fragmented internal sources (Gong, Slack, Pylon, Jira, Metronome, Salesforce, Notion) serving 30 Anyscale field engineers - Ray Data batch pipeline, recursive LLM summarization, dbt semantic layer, Claude agent skill, Streamlit app.

Cut weekly account-review prep from 2h to 20min across 16 tracked accounts (50+ person-hours saved weekly). Caught a book-wide 5–6x consumption-burndown underreporting bug; exec-summary format adopted by leadership for broader CS/AE rollout.

Shipped a production-grade vector-DB connector to Ray OSS with column-oriented batching, memory-safe sort+slice implementation, and a PyArrow hash-order bugfix.

Delivered faster write throughput and prevented silent data corruption, demonstrating Ray Data engineering work now part of open source and unblocking Notion's pipeline migration.

RCAd a 100%-timeout Ray service outage across Notion's EU production in under 3 hours by discovering a silent Python torch CUDA driver mismatch.

Outage restored in <3 hours; shipped a fail-fast torch.cuda.is_available() startup assertion as systemic mitigation.

Unblocked HeartFlow's CCTA clinical-imaging training across TBs of data by root-causing a prefetch-vs-consume OOM at 95.7% node memory. Shipped a dynamic-output-queue-size backpressure default upstream.

Replaced 20+ overnight manual restarts per training cycle with unattended sustained training - freed ML researchers and protected the Q1 renewal.

Prototyped a Ray Data reference pattern for autonomous-systems MCAP-to-tensor workloads, with on-the-fly H265 decoding and streaming reads from raw sensor data.

Reference design illustrating heterogeneous CPU-decode / GPU-train compute as an enablement artifact for customer evaluation - not a customer production deployment.

Built AI-powered video editor from scratch with AI agents as first-class citizens. Created 120+ automated videos for 5 pilot customers.

Reduced content creation time by 70%. Secured 1 paid pilot and 2 design partners.

Won Luma AI Hackathon as the only solo participant. Built tool that generates cinematic house tour videos from Zillow listings using AI.

Demonstrated end-to-end AI video generation pipeline producing 100+ videos from real estate data.

Placed 2nd at the E2B × Fireworks AI Hackathon (2024) by building a multi-agent documentation validator that extracts code from technical guides, defines success criteria, and runs the code in sandboxes to catch broken examples.

Generated 18 validation reports and surfaced broken or incomplete code in official OpenAI, Vite, and E2B documentation - caught ~25% of tutorial issues automatically. Featured by Fireworks AI.

Placed 3rd at the LangChain-sponsored Agents & Compound AI Systems Hackathon (2024) by building a multi-agent release-notes system.

Created a tool that transforms sparse release notes into rich documentation tailored for different audiences (CEO, developers, downstream teams) through coordinated AI agents.

Won 1st place at MongoDB GenAI Hackathon (2024) by building an AI-powered team matcher using MongoDB Hybrid Search (vector + keyword with Reciprocal Rank Fusion) on participant profiles, skills, and interests.

$2k in Fireworks AI credits awarded; dev cost stayed under $1 for the full prototype. Featured by Fireworks AI as the headline of their triple-hackathon placement article.

Currently building an AI-driven application designed to automate the categorization of personal financial transactions. This project involves using state-of-the-art Large Language Models (LLMs) to classify transactions based on descriptions and enriched context from Google search results.

The initiative aims to reduce manual categorization errors, enhance user experience by minimizing the need for manual input, and improve financial management efficiency.

Led 7-person team building LLM RAG system for life insurance policy processing. Built OCR ingestion pipeline using AWS Textract with LangChain integration.

Achieved 82% accuracy, reduced actuarial workflow dependencies by 64%, saving ~$5M annually.

Initiated and led the development of advanced machine learning models that successfully increased the customer win-back rate by 11%.

Enhanced customer retention strategies and set a new benchmark for predictive analytics in the telecommunications sector.

Managed a team of 9 data engineers and machine learning experts to implement cutting-edge ML solutions that enhanced core mining processes.

Reduced mining carbon footprint by 4%. Implemented MLOps best practices for model versioning and deployment, involving 18 models in production retrained monthly.

Designed and implemented a comprehensive org-wide quality solution that significantly enhanced the reliability of data products, driving business decisions.

Increased revenue by 18% within a single quarter through predictive modeling of customer fulfillment rates.

Led a team of 12 to overhaul SnapTravel's data platform, integrating advanced analytics that supported strategic business decisions.

Helped pivot the company from growth-focused strategies to profitability in just three months during the challenging first quarter of the COVID-19 pandemic.

Founded and led the development of a data intelligence platform integrating various data sources and accessed by thousands of users.