GATC Health
Senior AI/ML Engineer · GATC Health
Hybrid RAG + Agentic Research Platform
Built an enterprise AI research platform (think Perplexity, but for science) — used daily by lab teams to search the scientific literature, look up drug compounds, and explore hypotheses. Became the primary AI interface across the organization within weeks of launch.
Designed a hybrid search system that combines traditional keyword matching with modern semantic search — so it finds the right documents whether you phrase a query exactly or ask it conversationally. Achieves high recall across large chemical and biological literature corpora. (Technical: BM25 + dense embeddings, late fusion, cross-encoder reranking.)
Built a long-context memory system that lets the AI maintain coherence across multi-session research workflows — so scientists can return hours or days later and pick up exactly where they left off, with the AI retaining full context. Scales to millions of words without losing coherence.
Gave the AI the ability to autonomously use specialized tools — toxicity lookups, molecular property prediction, chemical structure searches, and live queries to scientific databases (ChEMBL, PubChem) — so it can answer complex drug discovery questions without a human in the loop.
Served neural network models on SageMaker with full MLOps — registry, CI/CD, canary rollouts, monitoring — and routed open-source models through LiteLLM and AWS Bedrock so newer models can be swapped in without rebuilding the application.
RAGBM25Dense EmbeddingsRerankingAgentic AILangGraphQwen 3.5SGLangSageMakerLiteLLMBedrockLIGHT Memory
PAL-3 Post-Hoc Clinical Analytics · VistaGen Therapeutics
Investigated why a large Phase 3 clinical trial for social anxiety didn't show the expected improvement over placebo. Combined patient health scores over time, site visit records, enrollment timing, and speech patterns measured from recorded therapy sessions to find out what actually happened.
Discovered that how patients spoke during visits — how much they talked, sentence length, total talk time — was a reliable predictor of who responded to placebo. This pointed to a habituation effect: the act of repeatedly attending clinic visits was itself reducing anxiety, making it harder for the drug to show a measurable difference.
Also explored how recruitment channels, site scheduling patterns, and patient vs. clinician communication styles influenced outcomes — delivering concrete recommendations to help the sponsor design more robust future trials.
Clinical NLPSpeech FeaturesLongitudinal AnalysisPlacebo ModelingPythonPandas
Graph Neural Networks for Molecular Property Prediction
Built ML models that predict drug compound properties from molecular structure — specifically whether a compound is toxic, how it gets absorbed and metabolized, and whether it can cross into the brain. Used graph neural networks (GNNs) that “read” a molecule as a graph — atoms as nodes, bonds as edges — rather than treating it as a flat list of numbers. F1 ≈ 0.90 · AUROC ≈ 0.92.
Handled the full data pipeline: sourcing compound data from public scientific databases (ChEMBL, PubChem), cleaning and deduplicating records, and addressing the common problem of heavily imbalanced datasets — where toxic compounds are rare, making them harder to learn from without special handling.
Deployed the models as low-latency GPU-accelerated APIs on Kubernetes— available to internal teams via well-documented endpoints, ready to integrate into broader drug discovery workflows.
GNNsPyTorch GeometricADMETBBB PredictionSMOTEKubernetesChEMBLPubChem
