I build backend systems for AI-assisted health tech and ONDC commerce.
Three years on Go, RAG pipelines, multi-agent orchestration — currently shipping the Clinical Copilot at UnitedForHer.
Selected work
UnitedForHer · 2025–present
A wrapper API over RAG + LLM for AI-assisted prescription generation. Doctors get grounded suggestions, edit inside dynamic templates with variable injection, and export as PDF.
18% latency drop · adopted end-to-end
Read the case study →Shayr Omnichannel · 2023–2025
Autonomous multi-agent orchestration over LangChain + OpenAI. A no-code UI lets non-engineers configure agent behaviour declaratively, replacing per-client integration work.
23% shorter dev cycles · 27% faster deploys
Shayr Omnichannel · 2023
Core protocol modules for ONDC commerce — data-mapping, API surface, and interoperability for retail and logistics partners. MVP delivered in the planned six-week window.
5k+ tx/day · 6-week MVP
Deep dive · Case study 01
The wrapper that hides the LLM — and the latency, and the hallucinations.
UnitedForHer · 2025–present · Go · OpenAI · RAG · React
Context
Problem
Approach
Architecture
The wrapper holds the contract.
Everything else is replaceable.
Decisions
A wrapper, not direct LLM calls.
One chokepoint for guardrails, rate limits, audit logging, and template variables. Every clinical artefact passes through the same lane.
Cached embeddings.
The same patient's records get hit many times per consult. Re-embedding on every call would have killed latency.
Render-time variable injection.
The LLM produces narrative with {{patient_name}} placeholders. Resolving at render time keeps templates human-editable.
Challenges
Hallucination safety.
Prescriptions are high-stakes. Strict prompt scaffolding, retrieval grounding for every clinical claim, and a final clinician-facing review before PDF export.
Latency.
Doctors won't wait. Embeddings cached; retrieval and the LLM call run in parallel where the prompt doesn't depend on retrieved context.
Doctor trust.
Adoption depended on the tool feeling like a faster pen, not a black box. Live preview + edit-then-export kept the clinician in the seat.
Outcome
Stack
Backend & Systems
Node.js
AI / LLM
RAG
Frontend
React
Data
MongoDB
About
“I draw the systems before I write them.”
— me, in a notebook somewhere
Started as an intern at Eunimart, then Shayr (Adya), where I went from API contributor to multi-agent system builder in three months. Now at UnitedForHer, building the Clinical Copilot.
Backend is home — Go, Spring Boot, message buses, microservices. The AI work followed naturally; LLMs are interesting because of the orchestration problems they create, not the model itself.
I draw — sketches, line studies, the occasional system diagram before I write any code. The drawing isn't a hobby alongside engineering; it's how I reason about systems before committing them to a repo.
The sketches in the margins of this site are the same ones in my notebook before each project starts.
Contact
— say hi