Boundaries that hold
Auth, RBAC, and tenant scoping aren't decorations — they're the contract. I ship fail-closed APIs so deleted data never stays retrievable.
Backend · Platform · Applied AI
I build thesystems layerbehind products that ship.
I turn messy requirements, auth boundaries, and production workflows into reliable systems people can actually use.
20+
PRs to production
77→90%
OCR accuracy lift
Working mode
Build · measure · iterate
01 / Systems mindset
Useful platforms start long before the model call.
Services, not scripts
I think in routers, repositories, and protocols — versioned contracts other teams can build on without breaking.
Observable AI backends
OCR, RAG, and eval pipelines with scores and traces. When models misbehave, I'd rather read the evidence than guess.
02 / Experience
Where I've been building.
The concise version. Projects carry the technical depth; this is the path that shaped it.
Jun 2026 — Present
Shipping into a multi-tenant document ingestion platform — Keycloak JWT + RBAC, document lifecycles, Marqo search, and self-hosted OCR/translation on H100s. 20+ PRs merged to production.
03 / Selected work
Projects that show how I think.
Less about model names. More about turning a problem into an interface, pipeline, or dependable decision surface.
/01Live productFull-stack RAG workspace
KAI
A deployed research desk that answers from retrieved PDF evidence — hybrid retrieval, reranking, and page-level citations.
Next.js
Qdrant
Gemini
RAG
Postgres
- Hybrid retrieval (dense + keyword + RRF) with Gemini reranking.
- Streaming answers with inspectable page-level citations.
- Multi-document chat persisted on Neon Postgres + Qdrant.
/02DeployedEnd-to-end booking platform
Airbnb Clone
A hiring-demo booking stack with layered services across explore, book, and host flows.
Next.js 15
FastAPI
SQLAlchemy
JWT
Alembic
- Nine domain entities with routers → services → repositories.
- JWT auth, Alembic migrations, wishlists, reviews, and host management.
- Next.js on Vercel + FastAPI on Render.
/03Production toolEngine evaluation harness
OCR Benchmark
A resumable benchmarking CLI used to pick production OCR engines on accuracy-versus-latency tradeoffs.
Python
vLLM
OCR
Eval
H100
- YAML model registry with per-engine runners.
- Scores ANLS / CER / WER plus a vision-LLM judge.
- Helped lift production OCR accuracy from 77% → 90%.
/04Research buildAI-image forensics CNN
FusionX-3
Triple-stream CNN for AI-generated image detection using RGB, DCT, and PRNU features.
PyTorch
Xception
CNN
Forensics
- Patch-wise Xception architecture across three forensic streams.
- Trained on 120,000 images.
- Reached 95% accuracy and 0.95 F1.
04 / Working stack
Tools change. Systems thinking compounds.
A practical stack for moving from raw requirements to something dependable in production.
01
Backend & Data
PythonTypeScriptFastAPIPostgreSQLSQLAlchemyRedisMinIO
02
Platform & Auth
KeycloakJWTRBACMulti-tenancyTemporalDockerLinux
03
AI / LLM Systems
RAGHybrid SearchQdrantMarqovLLMOCREval
04
Frontend & CS Core
ReactNext.jsTailwindDSADBMSSystem Design
Proof points
Production · 20+ PRs
Flywheel ingestion platform
Shipped auth, tenancy, lifecycle APIs, and OCR eval into a live multi-tenant document stack.
Search · 15K+ chunks
Semantic indexing
Tenant-scoped chunk search with Marqo, domain tagging, and a VectorStore protocol over direct vendor calls.
B.Tech ICE · 2023–present
NSUT, New Delhi
Instrumentation & Control at Netaji Subhas University of Technology — building systems on the side.