The gap between raw genomic data and actionable clinical insight is vast — and mostly unautomated.
My work sits precisely in that gap. I design and build AI-driven systems that transform high-dimensional biological data (single-cell transcriptomics, variant calls, clinical trial criteria) into structured, auditable intelligence. Whether that means orchestrating a multi-agent LLM pipeline to verify patient eligibility or engineering a high-performance variant detection tool in Python, the goal is the same: make the biology computable and the computation trustworthy.
| Status | Project | Focus |
|---|---|---|
| 🛠 Ongoing | ClinPilot | Expanding the FDA Guardrail agent with real-time ClinicalTrials.gov API integration |
| 🛠 Ongoing | Spatial Transcriptomics Pipeline | Adding cell-cell communication inference via CellChat |
| 🛠 Ongoing | MERN Clinical Dashboard | Building a patient-facing eligibility interface for ClinPilot |
How do you reliably match a real patient to the right clinical trial when the eligibility criteria span 40+ pages of regulatory language?
Problem Clinical trial eligibility verification is time-intensive, error-prone, and bottlenecked by human reviewers who must reconcile unstructured patient records against complex inclusion/exclusion criteria. A single missed criterion can disqualify a patient or expose a site to compliance risk.
Solution A 5-agent LLM pipeline where each agent owns a distinct epistemic role, running in sequence with structured handoffs:
| Agent | Role |
|---|---|
| Researcher | Retrieves and parses relevant trial criteria via RAG over ChromaDB |
| Advocate | Constructs the case for patient eligibility |
| Critic | Constructs the case against, stress-testing edge cases |
| Auditor | Reconciles the Advocate/Critic debate into a structured verdict |
| FDA Guardrail | Final pass for regulatory compliance and citation integrity |
The Granville Strategy underpins performance: a SQLite-based semantic caching layer intercepts repeated or near-duplicate queries before they reach the LLM, reducing latency and inference cost by short-circuiting redundant reasoning chains.
Tech Stack
Architecture: RAG → Multi-Agent Deliberation → Regulatory Guardrail → Cached Response
Caching: Granville Strategy (SQLite semantic cache) — reduces redundant LLM calls
Model: Llama 3.3 70B (instruction-tuned) via local inference
Problem Bulk RNA-seq averages over all cells in a tissue, obscuring the spatial organization of tumor, immune, and stromal compartments critical to understanding cancer progression.
Solution End-to-end spatial transcriptomics pipeline processing 10x Visium data: spot deconvolution, spatially-variable gene detection, and ligand-receptor interaction mapping to characterize the colon cancer tumor microenvironment at tissue resolution.
Tech Stack
Problem Gastric cancer subtypes are clinically heterogeneous, and bulk profiling fails to resolve the malignant, immune, and fibroblast populations driving treatment resistance.
Solution Single-cell RNA-seq pipeline from raw count matrices through clustering, cell-type annotation, differential expression, and trajectory inference — reconstructing the cellular landscape of gastric tumor samples.
Tech Stack
Problem Standard short-read variant callers under-perform on insertion/deletion detection in low-coverage or complex genomic regions, producing false-negative calls that matter clinically.
Solution High-performance Python implementation of a targeted INDEL detection pipeline built on GATK best-practices, with custom filtering logic and optimized I/O for large cohort processing.
Tech Stack
Problem Biological insights trapped in notebooks and scripts are invisible to clinicians, collaborators, and stakeholders who need them most.
Solution MERN-stack applications that surface genomic and clinical data through accessible, authenticated interfaces — from gene expression visualizers to role-based clinical dashboards with RESTful APIs and JWT-secured endpoints.
Tech Stack
| Domain | Tools & Methods |
|---|---|
| Single-Cell Genomics | Scanpy, Seurat, scRNA-seq clustering, trajectory inference |
| Spatial Transcriptomics | 10x Visium, Squidpy, spatially-variable gene analysis |
| Variant Analysis | GATK, INDEL detection, VCF filtering, cohort pipelines |
| Statistical Analysis | R/Bioconductor, differential expression, survival analysis |
| Data Engineering | Pandas, NumPy, high-performance Python, HPC/SLURM |
| Domain | Tools & Methods |
|---|---|
| Multi-Agent Systems | LangChain, custom agent graphs, deliberation frameworks |
| Retrieval-Augmented Generation | ChromaDB, vector embeddings, semantic search |
| LLM Integration | Llama 3.3 70B, OpenAI API, prompt engineering |
| Caching & Optimization | SQLite semantic cache (Granville Strategy), inference cost reduction |
| Regulatory AI | FDA-aware guardrail agents, citation integrity, audit trails |
| Domain | Tools & Methods |
|---|---|
| Frontend | React, JavaScript (ES6+), responsive UI |
| Backend | Node.js, Express, RESTful API design |
| Database | MongoDB, SQLite, data modeling |
| Auth & Security | JWT, role-based access control |
| DevOps | Git, GitHub Actions, Linux/Bash, Docker |
From: 25 February 2026 - To: 04 March 2026
Markdown 26 mins █████████████████████░░░░ 83.67 %
YAML 5 mins ████░░░░░░░░░░░░░░░░░░░░░ 16.33 %
"The most consequential code running today is the code that interprets biology."