Problem → approach → what it proves.

Synthetic electronic health records are sold as a privacy-safe, deletable stand-in for real patient data. But the field judges their quality almost entirely on associational tests: can a discriminator tell real from fake? do the distributions match? does a model trained on synthetic transfer to real? The flagship use of this data is the opposite kind of question, real-world evidence: what is a drug's effect on an outcome, adjusting for confounders. Nobody was checking whether a "realistic" twin preserves that.

I built a causal-fidelity benchmark: reproduce a full generator zoo (copula, GAN, VAE, tree-based, nearest-neighbour) on three national primary-care cohorts, then measure how far each one moves a confounding-adjusted treatment-effect estimate away from the real-data answer. The key device is an oracle floor (the error you'd get from a perfect reproduction against a fresh real sample), which turns vague "it's a bit off" gaps into hard multiples.

• Confounding-adjusted log-odds-ratios and g-computation, cross-checked with a doubly-robust estimator so the finding is a property of the generator, not the estimator.
• A synthetic ground-truth model with a known effect, to confirm the benchmark's ranking tracks true recovery (Spearman 0.96).
• A new neighbour-based method that reimposes the exact real per-feature statistics by construction while mixing multiple neighbours, so almost no record is copied (audited at under 2.5% exact copies).

Every generator distorted adjusted treatment effects 5–20× above the oracle floor, on all three cohorts. The sharpest finding: the Gaussian copula, which has provably perfect per-feature statistics, is the single worst generator for causal inference. And the discriminator score everyone gates on turned out not to predict causal error at all. Even my own method, the least-detectable privacy-safe generator in the zoo, still mis-estimates the magnitudes. That's the strongest form of the thesis: even the best associational twin fails the causal audit.

During drug approval, health authorities (FDA, EMA, PMDA and others) send waves of questions that experts must answer under time pressure, and every answer must be grounded, traceable, and auditable. Free-form generation is a non-starter: a hallucinated citation or a leaked cross-program document isn't a bad UX; it's a compliance failure. The hard question is how to make an LLM system genuinely useful when correctness and boundaries are non-negotiable.

A "compliance-native" system where the constraints are architecture, not afterthoughts: hybrid retrieval with reranking, deterministic structured outputs, grounded generation that flags any claim lacking evidence, mandatory human review, and metadata-enforced access boundaries. I framed the design as Human-Completing AI: the system extends what an expert can do without ever removing them from accountability.

• Hybrid retrieval: sparse + domain-adapted dense embeddings, fused and cross-encoder reranked.
• Expert-assignment as a constrained optimization, with expertise profiles learned from historical query/response pairs.
• An annotation-free evaluation that derives retrieval ground truth from citations in already-approved responses.

78% Precision@5 (up from 52% lexical / 61% dense), zero data-boundary violations across ~9,500 operations, 100% output consistency at temperature 0, and, in a 12-week pilot with 23 regulatory professionals, a 46% reduction in time-to-first-draft. I also published the honest negative result: response text is almost never reusable across programs (<2%); the value is in precedent discovery and evidence retrieval, not copying.

Choosing which disease to develop a drug for is one of the most expensive bets in science, made largely on static association and expert intuition. Real-world patient data holds a richer signal: population-scale timelines are, in effect, a corpus of natural experiments about what happens when a drug meets a disease. The challenge is reading a directional, time-resolved signal from them, and doing it in a way a scientist can actually trust.

φ is a representation-learning method over real-world EHR that produces a signed, lag-resolved interaction for every drug↔disease pair, then fuses it with two independent evidence layers (pharmacological mechanism and human genetics) into a single directional meta-path with full click-through provenance.

Before trusting φ on novel bets, I validated it the honest way. Unsupervised, it rediscovers established anti-diabetic biology: the sulfonylurea → KCNJ11/ABCC8 loop, GLP-1 → GLP1R, and SGLT2 → SLC5A2, all matching independent genetic-association evidence. The output is a ranked, explainable shortlist with an evidence ladder behind every candidate, running today on a multi-country data estate.

A startup building healthcare AI needs real production infrastructure, not notebooks: multi-country data ingestion, privacy-safe datasets, LLM pipelines that don't break, and a platform the whole product can stand on.

• A >5TB lakehouse on Databricks with a medallion architecture, plus large-scale processing pipelines spanning 6 countries.
• DataFlywheel: a synthetic-data generation engine over large multi-geography healthcare datasets, producing privacy-preserving data for downstream analytics and AI.
• Automated LLM content-generation and publishing pipelines that cut manual effort while holding quality and consistency.
• Backend services, LLM orchestration, and deployment: combining classical ML, LLMs, retrieval, and workflow automation into reliable production systems.

Working directly with the founding team, I set the platform architecture and infrastructure, made the key system-design and product calls, and translated business goals into technical roadmaps and milestone-based delivery, spanning AI research, platform engineering, and product strategy.

Tables, Figures & Listings (TFLs) are the statistical backbone of clinical study reports: enormous, exacting, and traditionally hand-built by medical writers with expensive external vendors filling the gap. Slow, costly, and error-prone in exactly the way regulated documents can't afford.

Rather than ask a model to write TFLs from scratch, I built a GenAI layer on top of a clean, validated analytics layer: generation grounded in structured statistical truth, with a human-in-the-loop workflow for the writers. The pattern that makes GenAI safe in a GxP setting: retrieve from a source of truth, generate structure around it, keep the expert in control.

Saved roughly $4.5M in external vendor spend and reportedly at least doubled medical-writer productivity across biologics and small-molecule programs, while keeping outputs auditable.

A global pharma supply network generates a flood of regulatory signal (inspection findings, audit observations, SOP deviations) scattered across formats and functions. Leadership needs to see systemic risk before an inspector does, but manually synthesizing it took weeks.

An agentic GenAI risk-intelligence framework that ingests heterogeneous regulatory documents, normalizes them, clusters recurring risk themes, and generates remediation roadmaps for global supply leadership, with the provenance to defend every conclusion. Presented publicly at the BioIT World Conference, Boston 2025.

Cut audit-preparation time by roughly 50%, compressing weeks of manual synthesis, and earned the BMS Science & Innovation Award, 2025.

Trial design decisions lock in enormous downstream cost and risk, but the tooling to model them was slow and resource-heavy: study setup alone took over a week.

Simulation of cost, risk, and timelines across design choices, and a ground-up rebuild of the trial-optimization tool to automate study ingestion and modeling.

Study ingestion dropped from 7–8 days to a few hours with half the resources, with estimated design-stage savings of $2–15M (a reported internal estimate).

Operators drowning in dashboards could see that a KPI moved but not why. Reconstructing the story by hand took analysts about a week, far too slow for daily decisions across a national network.

A copilot that decomposes KPI variance with causal inference, clusters the drivers over time, and turns the result into a plain-language narrative an operator can act on, with SHAP-based explanation so every attribution is defensible.

Collapsed a week-long reporting cycle to minutes, giving operators across 1,000+ centers same-day answers to "what changed and why."

Return-to-origin (shipments that never reach the customer and come back) is a major cost and a notoriously hard prediction problem across Indian e-commerce logistics.

A model that weaves geospatial signals, buyer-intent data, and operational levers to predict RTO risk per shipment, with the levers to actually reduce it, validated with major retail and D2C partners.

Reported to beat internal and industry benchmarks, and productized into a revenue-generating value-added service for major retail clients.

Target Malaria studies the population genetics of malaria-carrying mosquitoes to inform gene-drive strategy. Classifying the genetic origin of vector populations from high-dimensional SNP data demands both scale and scientific reproducibility.

Reproducible ML pipelines over millions of genomic features (automated quality control, linkage-disequilibrium pruning, and rare-allele detection) feeding interpretable classifiers for genetic-origin prediction.

A reproducible workflow for genetic-origin classification of Anopheles populations, done as a Visiting Research Scholar in the Burt Lab, Department of Life Sciences.

• Predicting severe outcomes in COVID-19 using only demographics, comorbidities and symptoms. American Journal of Emergency Medicine, Vol. 45, pp. 378–384 (2021).
• Demystifying Heart Failure with Mid-Range Ejection Fraction using machine learning. 48th Computing in Cardiology (CinC 2021).
• A machine-learning approach to detect congestive heart failure without ejection fraction. ICICA 2021 (abstract presentation).

• MIT COVID-19 Challenge — Winner (5,000+ global participants).
• MHRD Samadhan Challenge — Winner (800+ teams, Government of India).
• Microsoft Imagine Cup — India Finalist.