Buildingbeyondpossible.
I'm Sanat Mehrotra — a final-year CSE student building production RAG systems and backend services. Currently at Zeroto10.ai, engineering FastAPI endpoints and PostgreSQL data models behind an AI platform.
- Now
- Backend engineering intern, Zeroto10.ai
- Focus
- RAG, backend systems, applied AI
- Local
- IST Agra, IN
Selected Work
Upload a PDF, ask it anything, get an answer with the exact source it came from.
A seven-stage hybrid retrieval pipeline over text, tables and images. HNSW vector search on pgvector plus BM25 full-text, merged with Reciprocal Rank Fusion, MMR-diversified, then reranked by a Cohere cross-encoder. Ingestion is async via unstructured and PyMuPDF, with Gemini 2.5 Flash captioning every image so figures are searchable too.
- 7Retrieval stages
- ~15× lowervs GPT-4o cost
- text · tables · imagesModalities indexed
- FastAPI
- pgvector
- LangChain
- Next.js 14
- DeepSeek
Photograph an injured street animal, and the nearest verified NGO van is dispatched.
Gemini Vision scores injury severity from 1 to 10, PostGIS routes the report to the nearest verified NGO, and Socket.io drives live tracking across citizen, NGO, driver and admin views. A Telegram bot mirrors every action, so reporting a rescue never requires opening the web app.
- top 5,000 of 26,000+India Innovates 2026
- 1–10, automatedSeverity scoring
- PWA + TelegramEntry points
- Next.js 14
- Supabase PostGIS
- Gemini Vision
- Socket.io
Scored a design competition round that organisers were grading by eye.
Inception v3 generates similarity scores between a submission and the prompt reference, verifying compliance automatically. Wrapped in a Gradio interface and deployed to Hugging Face Spaces so the organising team could use it during the event with no setup.
- Prompt Rush 2.0Used in
- Inception v3Model
- HF SpacesDeployed on
- TensorFlow
- Gradio
- Hugging Face
- Python
What I do
I’d rather go deep on a narrow stack than claim everything. These are the areas where I can take something from an empty repo to something running.
At Zeroto10.ai I build and refactor the REST layer behind an AI platform — including reengineering the history retrieval endpoint into a central platform feature — and design the PostgreSQL schemas underneath it, with Alembic keeping every migration version-controlled and reversible.
- FastAPI
- PostgreSQL
- Alembic
- SQL
- JWT
QueryMind runs a seven-stage pipeline: HNSW vector search over pgvector alongside BM25 full-text, merged with Reciprocal Rank Fusion, diversified with MMR, then reranked by a Cohere cross-encoder before generation. Every answer comes back with the source it came from.
- pgvector
- BM25 + RRF
- Cohere Rerank
- LangChain
- Langfuse
Gemini Vision scoring stray-animal injury severity from a photo so the right NGO gets dispatched. Gemini Flash captioning images so they become searchable. Inception v3 scoring image similarity to judge a competition round. Each one replaces a manual step someone was actually doing.
- Gemini Vision
- TensorFlow
- PyTorch
- Hugging Face
- Prompt Engineering
Streaming chat UIs with cited source cards, a PWA with live tracking and push notifications over Socket.io, and a Telegram bot that can do everything the dashboard can — because the person reporting an injured animal on the street is not going to open a web app.
- Next.js 14
- TypeScript
- Tailwind
- Socket.io
- Supabase
Stack
Tools I reach for, and what I use them for
Languages
- Pythonprimary — backends, pipelines, ML
- SQLPostgres dialect, schema design
- TypeScriptevery front end I ship
- JavaScriptwhere TS would be overhead
- C++DSA and competitive programming
Backend & data
- FastAPIasync REST services, OpenAPI-first
- PostgreSQLschema design and query tuning
- pgvectorHNSW indexes for vector search
- Alembicversion-controlled migrations
- Supabasewith PostGIS for geospatial routing
- MongoDBdocument-shaped workloads
- Node.js / Expresswhen the stack is already JS
- Socket.iolive tracking and presence
- REST APIscontracts, versioning, error shapes
- JWTauth on QueryMind and PawAlert
- Firebaseauth and push on side projects
AI & retrieval
- LangChainorchestration and tool routing
- Embedding modelsGemini Embedding 2, Matryoshka 1536-dim
- Cohere Rerankcross-encoder reranking, v3.5
- Hugging Facemodels, datasets, Spaces deploys
- TensorFlowInception v3 image similarity
- PyTorchtraining and fine-tuning
- Langfusetracing and cost per LLM call
- Prompt engineeringmulti-query, MAP-REDUCE summarisation
- Gradiofast interfaces for ML tools
Front end & tooling
- Next.js 14App Router, streaming UI
- Reactsince first year
- Tailwind CSSon top of a token layer
- Gitrebase, not merge
- AzureFundamentals certified
Experience
- Jun 2026 — present
Backend Engineering Intern
Zeroto10.aiEngineering FastAPI endpoints for an AI platform, including reengineering history retrieval into a central platform feature. Designed the PostgreSQL schemas and Alembic migrations behind new onboarding features, and built the RAG block and web crawler powering the platform's agent.
- FastAPI
- PostgreSQL
- Alembic
- RAG
- Web crawler
- 2024 — present
Core Team Member
GDG on Campus, SRMRan Prompt Rush 2.0 and AutoCoder 2.0 at 250+ participants each, and independently secured two in-kind sponsors. On the technical organising team for the India AI Impact Summit 2026, and received the Event Organiser Award in March 2026.
- 500+ participants
- 2 sponsors
- Event Organiser Award
- ongoing
Volunteer
Robin Hood Army, NoidaCoordinating food distribution drives and community outreach.
- Community
Recognition
- India Innovates 2026Top 5,000 of 26,000+ · Bharat Mandapam
- HackIndia Web311th of 40 teams
- HackHacks15th of 42 · Cluster Innovation Centre, DU
- Event Organiser AwardGDG on Campus · Mar 2026
B.Tech, Computer Science & Engineering
SRM Institute of Science and Technology, Ghaziabad
Process
- 01
Frame the problem
Find the one job the system has to do. Everything that isn't that job gets cut before a line is written.
- 02
Prove the risky part
Build the piece most likely to fail first — retrieval quality, latency budget, model accuracy — as a throwaway.
- 03
Build the spine
Schema, contracts, and error paths before features. Migrations reversible, logs structured, failures loud.
- 04
Finish it properly
Motion, empty states, keyboard paths, and the mobile layout. The part people feel, and the part that usually gets skipped.