Clinical Review Agent
Retrieval-grounded summarization over patient-level records on self-hosted MedGemma, so data never leaves the network. Every statement carries a sentence-level citation, behind guardrails and a human review gate.
I build clinical data platforms and machine learning systems for life sciences, and still write the core code.
Ten years building clinical data platforms and machine learning systems for pharmaceutical and biotech teams. I still write the code that goes to production, not only the plan for it.
Most of it is retrieval systems and agents over patient-level clinical data. Every answer cites the record it came from, sits behind guardrails, and reaches a human reviewer before anyone acts on it. The stack is Python, R, SQL and PostgreSQL on AWS, shipped with Docker and CI/CD.
Underneath sits the clinical detail: CDISC SDTM and ADaM, legacy SAS rebuilt as validated R, risk-based quality monitoring and biomarker analytics, all under GxP. I also build in the open. I founded R4SUB, a seven-package ecosystem for submission readiness, and have authored ten R packages, nine of them on CRAN.
Ten years across clinical data products, engineering analytics and applied machine learning, most of it client facing and in audited environments.
Technical Lead, Clinical Data & Applied ML
TechWorksLab, embedded engineering for global pharma and biotech clients
Engineer, Scientific & Clinical Data Products
Appsilon, R and Shiny data products for pharmaceutical and research teams
Technical Lead, Engineering Analytics
AERODESIGN Solutions, full-stack analytical platforms and team leadership
Data Analyst
First Advantage, large-scale data validation, QC and statistical reporting
Leading in-house clinical application, analytics and ML delivery for sponsor clinical development, while staying hands-on in the code.
Retrieval, agents and tool use, evaluation sets, guardrails and human review gates; model APIs and custom MCP servers.
Open-weight inference (GGUF, llama.cpp, MedGemma); PyTorch, scikit-learn, transformers; vision ensembles (CoAtNet, ConvNeXt, DINOv2).
Expert R, production Shiny (modules, performance), CRAN package authorship, testthat, renv and reproducible environments.
CDISC SDTM and ADaM, Define-XML, controlled terminology, admiral and pharmaverse, RBQM and KRI frameworks.
SAS and SAS to R migration under validation, TFL shells and specifications, QC and double programming.
Python (FastAPI, plumber), REST API design, PostgreSQL, MongoDB, ETL and ELT, metadata-driven pipelines.
AWS, Azure, Docker, Linux, CI/CD (GitHub Actions, GitLab), Posit Connect; GxP and 21 CFR Part 11.
SAPs, survival, mixed-effects and Bayesian models; genomics, transcriptomics, proteomics; DESeq2, limma, edgeR, GSEA.
Clinical data platforms, machine learning and engineering tooling. Client work is shown anonymized. A cross-section, not an exhaustive list.
Retrieval-grounded summarization over patient-level records on self-hosted MedGemma, so data never leaves the network. Every statement carries a sentence-level citation, behind guardrails and a human review gate.
Scores and ranks trial sites and participants by risk, so oversight effort goes where the signal is. Indicator drill-down from study to site to subject, statistical thresholds, evidence tables and recalculation driving investigation.
Checks clinical datasets against CDISC SDTM, Define-XML and FDA conformance rules. Rules run deterministically and the model only explains each finding, giving a reason and a remediation path, which keeps it auditable.
Interactive biomarker visualization for exploratory analysis: spaghetti, box, volcano and heatmap plots with dynamic filtering and subgroup analysis. Re-engineered rendering and caching to hold interactivity on large datasets.
Demographics, disposition and treatment dashboards with patient timelines spanning adverse events, labs, medications, history and vital signs. Authored ADaM specifications and derivations with study-day logic and subgroup analysis.
Metadata-driven lineage platform linking Statistical Analysis Plans, analysis datasets and TLF outputs back to source metadata, with automated lineage tracking and traceability validation.
Medical image abnormality detection on the RSNA knee dataset with CoAtNet, ConvNeXt and DINOv2 ensembles. Antimicrobial peptide activity, MIC regression and hemolysis prediction; reasoning agents and evaluation for ARC-AGI.
End-to-end CDISC SDTM, ADaM and TLG reporting with automated Clinical Study Report generation. Cut report time from six weeks to two days.
R package suite converting raw clinical data into compliant SDTM and ADaM formats, with 200+ validation rules and automated compliance checks.
Self-hosted ML inference server with GPU acceleration, load balancing and API management. Sub-100ms inference at 10k+ concurrent requests.
Production RESTful API serving image-classification models. Responses under 200ms, JWT authentication, and horizontal scaling to 1000+ images concurrently.
50+ reusable Shiny UI components (modals, notifications, data grids, loading states) with a theming system. Adopted by 25+ applications.
CI/CD for R Shiny: quality checks, automated tests, Docker builds, security scans and multi-environment deploys with rollback. 99.5% success across 100+ releases.
Parameterized RMarkdown reporting wired into CI/CD for consistent, reproducible clinical and operational reports. 2000+ generated across 50+ report types.
R package combining web scraping with text mining and sentiment analysis. Adaptive rate limiting, topic modeling, and 92% sentiment accuracy across 1000+ sites.
Neural-network chatbot for customer support and FAQ automation. Context-aware replies, five languages, and routing of complex questions to specialists. 89% intent accuracy.
Secure corporate LMS for training and certification: courses, automated assessments, digital badges and enterprise SSO. 500+ learners.
Bridges between R and Python pairing R statistics with Python machine learning. Two-way data exchange with type conversion, validation and tuned memory management.
Packages and ecosystems I author and maintain. Nine are published on CRAN.
An open-source R ecosystem scoring clinical data readiness for regulatory submission. An ADaM-to-SDTM traceability engine, FMEA-based risk quantification and a Submission Confidence Index turn binary pass or fail checks into scores a reviewer can trace. Sole author and maintainer.
Local large language model inference inside R through GGUF model files and the llama.cpp engine, with completion and streaming and no cloud call. Built for environments where data cannot leave the network.
HTTP POST and query-parameter ingestion into Shiny applications as reactive values. Cross-platform and MIT licensed.
An interactive clinical dataset explorer for SDTM and ADaM data. Browse, filter and inspect analysis-ready datasets without leaving R.
Research Scholar
Indian Institute of Technology (IIT) Ropar
M.Tech, Artificial Intelligence
REVA University
B.Tech, Aerospace Engineering
Hindustan Aviation Academy
Open to clinical data and machine learning engineering, consulting, and collaboration. The fastest way to reach me is email.