Sanskar Nanegaonkar

Shipped a multi-tenant HR-policy RAG assistant for NTT DATA and two enterprise customers, indexing 1,000+ HR policies: FastAPI service with Qdrant per-tenant retrieval, cross-encoder reranking, OpenAI GPT synthesis, safety-aware policy-conflict detection, source-priority rules, and Langfuse audit traces. Owned multimodal SOP-generation and gap-analysis modules that convert desktop activity into operating procedures: multi-provider LLM pipeline (Gemini vision, GPT synthesis, Claude critic), LangGraph agentic multi-stage orchestration, Pydantic structured outputs, and write-review-rewrite loops; deployed across 3+ organizations with 100+ operators. Reduced manual transcript review per agent from a week to a few automated hours using a multi-agent RAG optimization system: planner / extractor / synthesizer chains, RAG over source corpora, human-feedback interpreter, variant regeneration router, and regression scoring against correctness thresholds. Shipped voice AI agents on Vapi and LiveKit for lead qualification, interviewing, and structured data collection; LLM dialogue policy with Pydantic tool calls and transcript-grounded evals matched human-agent pickup and conversion rates across 1,000+ calls. Replaced random spot-check QSC audits with continuous automated VLM compliance coverage, piloted live across 2 restaurant locations; designed a VLM-assisted auto-labeling pipeline that bootstrapped a proprietary training dataset in a kitchen domain with no open-source data or model coverage.

Built a multimodal depression-severity model using DeepFace facial emotions and librosa speech-prosody features (early fusion); reached 83% accuracy and extended to bipolar-vs-unipolar classification at 78% across 350+ patients. Productionized the clinical ML pipeline on AWS; collaborated with Keio on an FDA Breakthrough Device Designation application and Pre-Sub filings, and validated depression-severity predictions with Northwell Health and Hamamatsu across clinical cohorts. Cut therapy documentation time by 70% with an LLM system converting transcripts into Pydantic-validated SOAP notes; a clinician style-adaptation loop distilled edits into reusable per-clinician profiles.

Built video analytics pipelines across classroom sessions, live-streaming, and sales call recordings; extracted behavioral and linguistic signals (Python, OpenCV, scikit-learn) to distinguish high from low performers across 10,000+ recordings and generated structured coaching recommendations.