Apr 2025 – Present
Multi-Agent Workflows · Internal AI Research Platform
- Architected and shipped a production multi-agent AI research platform — an enterprise Perplexity for scientific teams — that became the primary AI interface for lab workflows within weeks of launch.
- Designed a hybrid sparse–dense retrieval pipeline (BM25/TF-IDF + dense embeddings, late fusion, cross-encoder reranking) for high recall on chemical and biological literature.
- Built a LIGHT-style memory subsystem that scales to millions of tokens of conversational history via episodic retrieval, structured working memory, and a compressed scratchpad — so multi-session research stays context-aware.
- Shipped domain agentic tools (toxicity lookup, molecular property prediction, structure normalization, ChEMBL & PubChem connectors) so agents can invoke specialized models and databases as callable actions.
- Served open-source models on SageMaker with SGLang; routed others through LiteLLM and AWS Bedrock for drop-in upgrades. Added evaluation harnesses and domain fine-tunes for alignment.
PAL-3 Post-Hoc Clinical Analytics · VistaGen Therapeutics
- Investigated a failed Phase 3 social anxiety trial by integrating longitudinal clinical outcomes, site operations data, enrollment timing, and turn-level speech features from recorded patient visits.
- Found 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 — with concrete recommendations for future trial design.
Graph Neural Networks for Molecular Property Prediction
- Designed and trained GNN architectures for toxicity, ADMET, and blood–brain barrier permeability prediction, choosing graphs over tree baselines to capture molecular topology — F1 ≈ 0.90, AUROC ≈ 0.92.
- Led end-to-end data curation from ChEMBL and PubChem: cleaning, deduplication, class-imbalance handling (SMOTE, weighted loss), and leakage-aware EDA.
- Deployed multi-modal ML models as low-latency GPU inference services on Kubernetes, consumed internally via documented APIs.