Vinu Kevin
Diesel

NEDL Labs · USA

• Architected a 4-stage pipeline turning dense healthcare-policy PDFs into executable SQL rules (LLM parsing, logic-tree flattening, rule assembly, DB ingestion) — replacing slow, error-prone manual policy interpretation so rules are queryable the moment a document lands.
• Built a 5-layer validation framework that cross-checks every LLM-extracted rule against the source text via fuzzy matching and forward/reverse coverage analysis, catching hallucinated or missing logic before it ships and shielding downstream billing & coverage decisions from silent errors.
• Developed a policy-domain classifier with multi-tier scoring, circuit breakers, and confidence gates that auto-routes policies across 10+ medical specialties — letting the pipeline scale to new domains with zero manual triage.

NEDL Labs · USA

• Built a PDF-to-structured-data pipeline for casino gaming reports using Detectron2, EasyOCR, and table-recognition models in a 4-container Docker service — turning slow, manual report data-entry into an automated extraction step.
• Implemented automatic column matching with 384-dim embeddings in PostgreSQL (pgvector), unifying data from 39+ casino properties into one standardized schema — eliminating hand-mapping and unlocking cross-property analytics.

Hewlett Packard Enterprise · Singapore

• Built a custom posture dataset via web scraping and MoveNet keypoint extraction, converting joints into normalized vectors — the labeled foundation that made reliable classification possible.
• Trained a multi-layer CNN with dropout and hyperparameter tuning to 98% accuracy — precise enough to flag unsafe postures in real time for workplace-ergonomics and safety monitoring.
• Deployed on Azure Databricks with Spark batch-inference pipelines and live dashboards, turning raw model output into a real-time monitoring tool — presented to faculty and industry professionals.