Clinical evidence-appraisal copilot (Elyx, 2025)
Multi-agent system that synthesizes findings from top medical literature, clinical guidelines, and trusted sources into citation-backed answers for care teams.
Ashish Kumar / Applied AI Engineer
Healthcare AI · Agentic Systems · LLM Infrastructure
Currently building production AI at Elyx — agentic clinical workflows, OCR-to-FHIR pipelines, and evaluation harnesses on AWS and GCP.
Why this work
Most AI systems fail not because the models are weak, but because the surrounding infrastructure isn't ready for production.
I build trustworthy AI systems for high-stakes domains — healthcare, scientific discovery, anywhere accuracy and traceability matter as much as capability. My work spans agents, retrieval, OCR-to-FHIR pipelines, healthcare interoperability, LLM evaluation, and the cloud infra that ships them. The goal is simple: help clinicians spend less time searching for information and more time making informed decisions.
Recent work
Where I've built.
2025 → Present
Elyx
AI Engineer
Production AI for healthcare. Shipped a clinical research and evidence-appraisal copilot, OCR-to-FHIR pipelines, and agentic workflows for longitudinal patient review. Operate the cloud infra (AWS + GCP, IaC) and add evaluation, tracing, and human-in-the-loop guardrails.
Aug 2024 → Jul 2025
Instel LLC
Generative AI Engineer
Biomedical research and literature intelligence. Multi-agent workflows for information discovery and synthesis; deep-research workflows for efficient report generation.
Dec 2023 → Aug 2024
Harvard IACS
Generative AI Engineer
Built AstroRAG over the NASA ADS corpus. End-to-end LLMOps pipeline with the Harvard–NASA ADS team. Fine-tuned Mistral 7B for abstract optimization and impactful titles.
May 2024 → Jul 2024
DeNA
Data Science Specialist (PPO)
User analytics and behavioral modeling for the Pococha live-streaming platform. Bot detection via user–IP interaction graphs (NetworkX); spending-based cohorts for retention.
2023 → 2025
IIT Kharagpur
Researcher · Co-Founder · TA
Biomedical IR research: fine-tuned BERT for ICD-10 retrieval over PubMed, DailyMed, and clinical-trial data; evaluated with Precision@20 and MAP. Co-founded SPARK4AI. Teaching Assistant for an AI/ML course.
Capabilities
Four overlapping disciplines.
Healthcare AI
FHIR-native pipelines, OCR over clinical documents, clinical literature retrieval, evidence appraisal, EMR-resident agents.
FHIR · OCR-to-FHIR · clinical retrieval · EMR · evidence appraisal
Agentic Systems
Multi-agent orchestration, tool use, planning loops, grounded retrieval with citation, evaluation harnesses.
LangGraph · multi-agent · tool use · grounded retrieval · Langfuse
LLM Infrastructure
Production AI services on AWS and GCP. Reproducible deploys via IaC, container-native runtime, observability and tracing in healthcare-critical environments.
AWS · GCP · Terraform · Docker · CI/CD · evaluation · observability
Engineering
Async backends, FastAPI services at scale, TypeScript frontends for clinical interfaces, evaluation tooling.
Python · FastAPI · PyTorch · TypeScript · React · SQL
Selected work
A handful of things I've shipped.
OCR-to-FHIR pipelines (Elyx, 2025)
Convert scanned reports and unstructured clinical documents into structured, queryable healthcare data so PDFs become first-class citizens in EMR workflows.
AstroRAG (Harvard IACS, 2024)
Context-aware retrieval over the NASA ADS astronomy corpus. Mistral 7B fine-tuned for abstract optimization and title generation; end-to-end LLMOps pipeline.
Recognition
Markers along the way.
- Kaggle Competition Master
- AI4ICPS Hackathon Winner
- Geoffrey Hinton Fellowship
- IIT Kharagpur · Dual Degree (M.Tech AI/ML + B.Tech Biotech)
- DeNA · PPO
PUBLICATION
An AI-driven system for enhancing Astronomical Research workflows
Latest writing
Notes from the workbench.
Contact
Open to work that matters.
Open to senior IC roles building AI for healthcare, scientific discovery, or other high-stakes domains.
