Priyanshu Agarwal

Stock Finder is a hyperlocal marketplace that lets buyers in India find which nearby kirana and retail shops actually have an item in stock, instead of calling around or traveling shop to shop. Shop owners maintain a lightweight catalog; buyers search by item and see ranked results by distance, with a direct link to the owner.

I built the product end to end. On the back end, I designed a domain-split Django 5 monolith (auth/OTP, shops, inventory, leads) that exposes a REST API, with PostGIS for geospatial proximity search and a hybrid Postgres search layer that combines trigram similarity and full-text search so misspelled or partial product names still match. Redis caches nearby-shop lookups, and Celery handles asynchronous work such as image variant generation and stale-inventory sweeps. Cross-app coupling is mediated through Django signals and Redis pub/sub, which surface as real-time updates to the client over SSE.

On the front end, I built a React 19 + TypeScript SPA with Chakra UI, TanStack Query, and MapLibre GL for map-based shop discovery. I also containerized the stack and wrote the Helm charts for Kubernetes deployment.

Atlas is a full-stack geospatial and AI workspace I designed and built end-to-end. The back end is an event-driven, modular Django monolith with 10+ domain apps on PostgreSQL/PostGIS, Celery workers, Redis pub/sub, SeaweedFS S3-compatible object storage, and Kubernetes deployment. Users upload large GeoTIFF rasters; Celery converts them to Cloud Optimized GeoTIFFs, and the API serves XYZ map tiles rendered on demand with rio-tiler. Django signals publish to Redis and stream to the browser over Server-Sent Events, so clients stay in sync without polling. An agent module built on LangGraph and LangChain (local Ollama plus OpenAI models) drives chat over WebSockets with pgvector-backed RAG.

The React 19, TypeScript, and Vite front end is feature-sliced, using React Query for all server state and MobX confined to the GIS domain. The map subsystem follows a ports-and-adapters design, so MapLibre GL and Terra Draw stay behind engine interfaces and never leak into application code. Chakra UI powers a token-based design system, and quality is ensured by Playwright E2E tests, Vitest/Storybook component tests, and k6 load tests that gate API changes.

An AI system that reads an entire book series and answers questions with page-exact citations, without ever making up facts. Most chat-with-PDF tools give plausible answers that they can't back up. In Traverse, every claim links to the page that proves it. I designed and built it end to end as the sole engineer.

Users upload novels as PDFs and get a knowledge graph of every character, every relationship between them, and the pages that support each one. Questions in plain English get streamed, cited answers, and are limited to the reader's current chapter, so nothing is spoiled. When the AI is unsure, it stops and asks a human instead of guessing.

What I built:
• PDF pipeline: layout-aware parsing (Docling), page-level provenance, embeddings in Postgres/pgvector, run as Celery jobs
• LLM extraction that links each person's many names and saves only evidence-backed relationships to Neo4j
• LangGraph agent that chooses between graph queries and hybrid search with reranking, plus a human review queue
• Private, fully local inference (Qwen3-8B on vLLM) with Langfuse tracing and an evaluation harness
• React/TypeScript front end with an interactive graph explorer

The same approach fits contracts, research, and knowledge bases.