Director, AI Scientist at IQVIA · Ph.D. ECE, Stony Brook · M.B.A., Boston University · IEEE Senior Member

I lead an AI/ML portfolio and a team that builds it — and I still write the code. Most of what I ship is behind an enterprise firewall, so the repositories here are the ones I built on my own time to keep the hands-on half honest: real deployments, real users, real evaluation numbers including the ones that are not flattering.

Live now — indhive.com · cloverailab.com · pumpkinsolve.com

LinkedIn · Google Scholar · [email protected] · Malvern, PA

Applied AI in regulated, evidence-heavy domains — pharmaceutical commercial analytics, clinical development, and regulatory workflow. The recurring problem in this space is not model quality. It is whether you can defend the output: where each value came from, what the system refused to decide, and how you know it is right.

That constraint shows up in every project below. Three of them are deployed and open in a browser right now, so you can check.

indhive.com · source · Python · TypeScript · Cloudflare Workers

Sponsor, product, protocol and plan records go in; a deterministic pipeline normalises them into one canonical record and maps that into FDA Module 1 artifacts — Form 1571, the 1.20 General Investigational Plan, a Module 1 gap analysis, a drafted cover letter.

Two design decisions carry the project:

• Every value keeps the source record it came from. Provenance is not a logging afterthought; it is part of the data structure.
• Where two source records disagree, the field is left empty and flagged. The system does not choose. In a regulatory filing, a plausible guess is worse than a blank.

The cover letter is drafted from an approved-fact whitelist and then checked against those facts for unsupported claims. The assistant on the site never states a case fact that did not come from a tool call, never resolves a conflict, and never claims filing readiness.

How I know it works: a clinical partner built a real-shaped input package for a programme of his own and sent his own answers with it — a hand-filled Form 1571, a ten-entry gaps log, and a traceability matrix declaring 23 numbers that must agree across his dossiers. The pipeline ran on his input and never saw his answers. Four comparison views put the two side by side; four of his ten gaps came back independently.

Every case is fictional and nothing here is submittable to FDA. It is a proof of concept, not regulatory software — and the README says so before you ask.

cloverailab.com · source · TypeScript · Cloudflare Workers · Claude

An essay reviewer built on LLM-as-judge with an explicit ETS rubric and official scored anchors, plus a writer agent. No fine-tuning — 30 ground-truth essays cannot train a scoring model, and pretending otherwise is the usual failure mode in this category.

Measured, not asserted — leave-one-out over the gold set:

The negative bias is real: the reviewer scores slightly harsher than human raters. It is in the repository along with the worst misses, because a benchmark you can only pass is not a benchmark.

The part I am most pleased with is a data bug. Seven of the eighteen official Issue essays in the public corpus had rater commentary spliced into the essay text — one "1,039-word" score-6 response is 775 words of essay and 264 words of a grader saying the response demonstrates facility with the conventions of standard written English. Left alone, the highest-scoring anchors would have taught the model that sounding like a grader is what a 6 looks like. The extractor re-derives them from the source PDFs and separates the two, and there is an assertion in the validator that fails if it ever regresses.

Python · Pydantic · structured output

Scores a job description against a fixed ontology of 44 life-science domains, 1–10 with a written reason per score. The scoring is the easy part; the reason this one is here is what surrounds it.

The calibration is an offset, and it is written down. A clinical panel hand-scored a seed set to establish the baseline. Aligning the model to it revealed that the panel judged relevance about one band more inclusively than the model does by default — where a domain sits in the macro background of a senior role, the panel called it moderately relevant and the model called it weak. That offset is stated in plain language in the rubric, along with the specific disagreements found during calibration, rather than being absorbed invisibly from a few-shot set. It is a thumb on the scale, deliberately, and a reader can disagree with it.

Then it was checked. Two independent scoring runs were compared across ~1,000 jobs × 44 domains. The first version of that analysis reported band agreement alone, which flatters when scores cluster; it now reports quadratic weighted kappa alongside a signed mean difference, so "these two disagree" is separable from "one sits a point above the other" — the second being a calibration offset, and fixable.

pumpkinsolve.com · source private · Next.js · Cloudflare Workers · Gemini

Somewhere to bring a half-formed practical question — about a workflow, a piece of research, an application — and have an assistant help frame it, offer a first pass, and compare directions, in a session no one else can see. It sells nothing and takes on no work; the point is to watch what people actually ask when the question is still unformed, because that is where the useful tools turn out to be hiding.

The next repository is the evaluation harness built to keep it honest.

JavaScript · zero dependencies

The harness behind the assistant above, generalised. Three ideas, each aimed at a failure a turn-by-turn review cannot catch:

Rules before the judge. "Turn 3 asked four questions" is a fact; handing it to a model invites an opinion about it. The rule layer costs nothing and has no variance, so it runs on every commit — including the check that the agent never asks a member of the public for a password, a card number, or medical records. When a rule and the judge disagree, the rule is right.

Convergence is a property of the transcript. An agent can be correct on every single turn and still never say "I have enough — here is the next step." Score it turn by turn and every turn passes while the conversation fails. So convergence is measured over the whole transcript, with a declared window: converging on turn one is as much a failure as never converging.

Flapping is its own verdict. Every case runs N times. A case that passes two runs out of three is not a passing case — it will fail in front of a user eventually, and it needs a different fix from one that always fails. It gets its own bucket instead of being rounded into a pass rate. There is a test asserting that a single run cannot tell flapping from passing.

42 peer-reviewed publications · 858 citations · h-index 15 · 4 granted U.S. patents, 2 pending

Recent work in IEEE TPAMI, NeurIPS, and CVPR (2024–2026), on adversarial robustness and distributed learning. Ph.D. dissertation: Distributed Bayesian Learning in Multi-Agent Systems (advisor: Petar M. Djurić).

Primary inventor on patented HCP-level promotion response and multichannel media attribution methods now used in commercial pharmaceutical measurement.

→ Google Scholar

→ Research collaborations — an index of the published work I contribute to with academic groups and where the code lives, with full author lists and BibTeX. The linked repositories belong to the students and researchers who did the experimental work; they are linked there, not forked here.

Includes a seven-paper collaboration on transferable adversarial attacks running since 2023 — ICCV 2023, CVPR 2024, NeurIPS 2024, IEEE TPAMI 2026 — and notes plainly which lines of work have no published code, and why.

At IQVIA I hold technical and delivery leadership for an AI/ML portfolio in pharmaceutical commercial analytics, with a team spanning data science, ML engineering, data engineering and software engineering. The work I am proudest of there is the unglamorous kind:

• Violet — the enterprise AI platform the organization's data science teams build on; 17 applications, 50+ active builders, consolidated from several overlapping stacks.
• MCP Factory — a governed registry for MCP servers and AI skills serving a ~500-person business unit: submission templates, an automated review step, a human approval gate, and active curation to retire duplicate capabilities.
• A responsible-AI operating model — risk tiering, approval gates, evaluation thresholds, human-in-the-loop rules and auditability — with a remit covering every AI platform the unit uses, including ones other teams built. Run in standing partnership with information security, data privacy, and legal.
• An evaluation stack that decides whether things ship: benchmark sets drawn from historically scored data, release thresholds, expert review fed back into the agent, and LLM-as-judge with multi-agent cross-scoring.

None of that is on GitHub. The two projects above are how I keep the skill.

Languages Python · TypeScript · SQL AI Claude / Anthropic API · Gemini · LangGraph · MCP · DSPy · Bedrock · evaluation harnesses Data & platform Databricks · Snowflake · Spark · Azure · AWS · Cloudflare Workers Serving FastAPI · Streamlit · React