Vineel Valmiki

CarParts.com Machine Learning Engineer and Project Lead · Contract

Leading a three-person engagement that replaces assumption-driven stocking with forecast-backed sourcing across roughly 1.18M live SKUs, reporting to the CTO weekly.

• Two-stage LightGBM pipeline predicting whether a part will sell and how much: 0.917 ROC-AUC and 0.772 precision@1,000 in rolling validation.
• Cut aggregate under-forecasting from 21.57% to a 3.77% mean absolute gap.
• Spark ETL across thousands of tables, including one 600 GB source, into a 48-month quarterly panel of 6,754 categories.
• Caught a rolling feature leaking future information into training, rebuilt 20+ features, and reconstructed the folds.

Software Product @ Cal VP of Consulting

Source and close paid client engagements worth $15K to $35K each, turn vague asks into scoped statements of work, and manage delivery. I lead a 15-person organization and run workshops that get new members ready for billable work.

SpaceX Technical Product Manager · Contract

Co-led a tiered global reseller partner program for the Starlink Commercial Team, covering 200+ partners across 150+ countries.

• Built a composite scoring model (revenue 45%, verified services 30%, activation 25%) indexed by region so partners compete on penetration, not raw dollars.
• Designed an evidence-based verification pipeline with AWS Textract and Grok Vision feeding LLM scoring, with low-confidence cases routed to humans.
• Presented the architecture and phased rollout to Starlink leadership.

Granite Construction AI/ML Engineer · Contract

Built bid-pricing models across 3,900+ Caltrans contracts and 5,900+ bidder records: quantile regression for bid ranges, win-probability models, and a plain-English retrieval advisor (FAISS, Flask) so executives could question the model. Demoed live to the Director of Technology.

Vectorial AI Machine Learning Systems Engineer Intern

Designed evaluation case studies for a synthetic-audience product and raised measured synthetic-to-real alignment from 80% to roughly 90% across 15+ customer segments. Built the scraping and async worker infrastructure behind it (100+ concurrent tasks).