Hi, 

Hybrid RAG + Agentic Research Platform

Architected and shipped a production-grade internal AI research platform — functioning as an enterprise Perplexity for scientific teams — that became the primary AI interface for lab workflows, significantly boosting researcher productivity and usage across the organization.

Designed a hybrid sparse–dense retrieval pipeline combining BM25/TF-IDF with dense embeddings, late fusion, and cross-encoder reranking to achieve high recall on scientific literature across large chemical and biological corpora.

Built a LIGHT-style memory subsystem that scales to millions of tokens of conversational history using episodic retrieval, structured working memory, and a compressed scratchpad — enabling persistent, context-aware interactions across long multi-session research workflows.

Developed domain-specific agentic tools: toxicity lookup, molecular property prediction, structure normalization, and external database connectors (ChEMBL, PubChem) — allowing the LLM to autonomously invoke specialized ML models and data sources as callable tool actions.

Self-hosted Qwen 2.5 27B on AWS SageMaker using SGLang for high-throughput, low-latency inference. Modular architecture supports drop-in upgrades to larger or newer models. Implemented evaluation harnesses and fine-tuned instruct models for domain-specific alignment.

PAL-3 Post-Hoc Clinical Analytics · VistaGen Therapeutics

Investigated a failed Phase 3 social anxiety trial (negative primary endpoint) by integrating longitudinal clinical outcomes, site operations data, enrollment timing, and turn-level speech features extracted from recorded patient visits.

Discovered that subject speech dynamics during treatment visits — conversation share, utterance length, total talk time — carried a consistent, leakage-aware predictive signal for placebo response. Developed a habituation hypothesis explaining unexpected placebo performance.

Explored recruitment channel effects, site-level calendar drift, psychometric symptom structure, and clinician vs. patient vocal dynamics as outcome moderators — providing actionable recommendations for future trial design.

Financial Risk & Pricing ML · DriveTime · Carvana

Led development of production ML systems for financial risk and pricing products used by enterprise customers including DriveTime and Carvana — directly impacting underwriting decisions at scale across hundreds of thousands of auto loan applications.

Designed, trained, and deployed predictive models for risk-adjusted APR, underwriting, LTV, depreciation curves, and delinquency prediction — achieving F1 scores close to 0.90 across multiple use cases with rigorous backtesting and holdout validation.

Owned customer data pipelines and feature engineering workflows, partnering closely with client data teams on EDA, labeling strategy, feature importance analysis, and model validation — including fairness and bias audits for regulatory alignment.

ML Inference Platform · AWS Kubernetes

Led the team to build a scalable microservices inference platform on AWS Kubernetes with full observability — ELK stack, structured logging, latency monitoring, and auto-scaling to handle burst inference workloads across multiple model families simultaneously.

Authored custom Docker images with CUDA toolkit support and container runtime integration for GPU inference, enabling hardware-accelerated model serving for deep learning workloads with consistent performance across environments.

LMS Rebuild [Case Study]

Led the full-stack rebuild of a large-scale Learning Management System serving millions of users — achieving approximately 75% improvement in frontend performance scores through architecture optimization, code splitting, and caching strategies tailored to 2G mobile demographics across India.

ML Product: “Shorts” — Micro-Learning Feed

Conceived and built an ML-powered micro-learning product — a TikTok-style short-form feed — end-to-end. Implemented the SM-2 (SuperMemo-2) spaced-repetition algorithm enhanced with a Feed-Forward Neural Network to personalize content sequencing and retention scheduling per user, adapting review intervals based on individual performance signals.

Yahoo Ad.com — Self-Serve SMB Ad Platform

Contributed to Yahoo Ad.com, a self-serve advertising platform enabling small and medium businesses to create and manage ad campaigns across the Yahoo network with minimal setup friction.

Built ML-assisted onboarding flows that used basic business inputs — website URL, name, category — to auto-generate initial campaign configurations, measurably reducing time-to-first-campaign for new advertisers.