Pulkit Saraf

I'm an engineer who lives at the intersection of AI and systems. I like building autonomous agents, LLM and SLM backed services, and the backend that keeps them fast and reliable at scale. Mostly, I just enjoy making complex systems feel simple.

I care about the details that make software actually hold up in production, but I don't take myself too seriously. You'll usually find me deep in a new project or a hackathon, learning something on the fly, shipping it, and probably fixing it later with a coffee in hand.

Lately I've been obsessed with agent orchestration and the small design choices that decide whether a system scales or falls over. If it involves reasoning models, weird distributed edge cases, or making something feel effortless, I'm in.

Professional Experience

February 2026 – Present · Remote

November 2025 – December 2025 · Chennai, India

June 2025 – October 2025 · Remote

March 2025 · Remote

Technical Contributions

I build autonomous agent systems end to end, from an X (Twitter) agent that analyzes posts and recommends content to client-facing model inference APIs at Rabbit AI, where I worked on small language models and improved model accuracy by 20%.

Opsuna is a generative UI platform I engineered that compiles natural language into production-ready React components in real time. It combines LLM NLP pipelines, Zod schema inference, and WebSocket streaming, with three agents over pgvector semantic search sustaining a 96% cache hit rate across 1.3 billion tokens.

Blink AI is a multi-tool desktop assistant with over 240 tool integrations, real-time Deepgram speech recognition, screen OCR, and mobile-to-desktop command relay, built on a plugin architecture with zero-config external API integration.

agmonverse is an ML optimization engine running predictive modeling and anomaly detection with sub-second inference over live data streams. It won Monad Blitz against more than 50 teams and shipped to production.

ferrokey is an embedded key-value store I built from scratch in Rust, with a custom B-tree index, a fixed 4096-byte page format, and a write-ahead log with CRC32 checksums and fsync guarantees. Crash recovery replays the log on open, with zero runtime dependencies.

In open source, I have contributed 21+ merged PRs across CNCF projects like Jaeger, Meshery, and OpenFGA, spanning gRPC error handling, storage internals, and CVE cleanup. I have applied to three LFX 2026 mentorships and I am contributing to Google Summer of Code 2026.