Disha Faujdar — Software Engineer & AI Practitioner. Portfolio and projects.

Built an AI-powered health tracking mobile app with Expo, Supabase, and PostgreSQL. Improved screen load times by ~35% via soft reloading, client-side caching, and incremental state updates. Enhanced in-app search with prefix-based and fuzzy matching, improving successful search results by ~45%. Tech: Expo, Node.js, Prisma.

Seren investigates production incidents autonomously — receives webhooks from PagerDuty, Sentry, and Alertmanager, runs a ReAct agent loop querying Prometheus, Loki, and Git, and delivers a root cause verdict with confidence score before the on-call engineer opens their first tab. Built with FastAPI, ARQ, SQLAlchemy, Qdrant for semantic incident memory, Neo4j for service dependency graphs, and a handwritten ReAct loop with loop detection, tool timeouts, and Redis pub/sub for live reasoning trace streaming.

Built a platform-agnostic synchronized music listening app where users drop into a shared room via link and listen together in real-time — no account required. Build a sync engine: a drift-correction system that computes true playback position accounting for network latency, then applies soft corrections via playback rate manipulation rather than jarring hard seeks. Unified a single IPlayer interface across YouTube, Spotify, and SoundCloud SDKs, each with wildly different APIs, into one coherent abstraction. Built on Next.js with WebSocket server since App Router doesn't support persistent connections natively.

Built a persistent memory layer for autoresearch by karpathy — an autonomous ML research agent that runs hundreds of experiments overnight. The problem: the agent rediscovers the same dead ends every run. No memory of what it already tried.The architecture is directly inspired by CoALA (Sumers et al. 2023) — episodic memory for autonomous research agents.

Fillit is a job application autofill system: a Next.js dashboard to manage your profile and resumes, a Chrome extension that fills application forms using a native setter strategy cascade, and a Node/Express API with PostgreSQL.

WikiMind — full-stack RAG system where users upload PDFs and chat with their documents. Built adaptive retrieval hybrid BM25 + reranking based on corpus size, hierarchical chunking, streaming responses, used async pipelines, redis and bull.js for queuing and background jobs, and end-to-end observability via Langfuse.

AI Research Agent that accepts a topic, runs a small research workflow, and returns a structured result. Tech: TypeScript, Redis, Bull.js, Express, PostgreSQL.

Conservative decision system matching developers with realistic open-source contribution opportunities. Analyzes proven work history and matches with contributor-friendly repos. Tech: Python, Typer, Rich, GitHub API, Sentence Transformers, FAISS, LLMs.