Rohan Sehgal
I'm an Engineering Fellow at Tools for Humanity, working on World ID.
Previously, I was employee #3 and a founding engineer at Broccoli. I helped scale the company from $0 to $10M+ ARR and built systems powering millions of AI voice calls for home-service businesses.
WORK EXPERIENCE
Tools for Humanity
June 2026 — Present
Engineering Fellow
Building identity systems on the World ID team.
Broccoli.com (YC W22)
January 2025 — May 2026
Founding Engineer
Employee #3. Helped scale the company from $0 to $10M+ ARR. Built systems powering millions of AI voice calls for home-service businesses. Backed by Y Combinator and Khosla Ventures.
Amazon Robotics
September 2024 — December 2024
Software Development Co-op
Building automated AMI creation workflow to standardize EC2 config, installing Docker, NVIDIA drivers, and ECS agent config, cutting setup time and ensuring consistent, high-performance execution for GPU-accelerated tasks. Reduced total runtime from 45+ to 10 minutes, while improving stability and preservation of artifacts. Developed Hydra CLI automated integration tests to ensure stability for large-scale synthetic data generation.
Salesforce
June 2024 — September 2024
Software Engineering Intern
Worked on IAM Copilot to automate and optimize internal IAM support using a GenAI-driven backend system Utilized Python and XSOAR to support additional features through agents Cross team collaboration effort with IAM team, boosting real-time user engagement and speeding access to support
PROJECTS
Projects, experiments, and systems I've built. Each item opens the work itself.
Rust EAGLE-3 — draft adaptation
Adapted an existing speculative draft for frozen Qwen3-8B. Measured 15.6% lower Rust latency versus the upstream draft on a small held-out workload, with seed replication, matched training controls, and explicit serving regressions. Includes reproducible code, raw evidence, and six checkpoint cards.
ctf-lite — RL training environment
Turned a capture-the-flag crypto eval into a GRPO training environment for Qwen3-4B on prime-rl. Kept an honest null result (no significant lift at n=16 rollouts), diagnosed it to a difficulty cliff that starved GRPO of gradient, then built a procedural difficulty generator that splits each task across decoupled sub-knobs to turn a cliff into a climbable curriculum. Published as a repo, a wandb report, HF weights, and a Prime Hub environment.
Reverse-LLM
A language model that reads forward-English questions and answers in character-reversed English, generated tail-first. Custom BPE trained on a mixed forward + reversed corpus, direction tokens for mode control; standard decoder-only transformer, forked from nanochat.
Website Cloning Agent
Agentic system that turns any URL into a dev-ready multi-page Next.js clone. Playwright crawls and inspects the source site at desktop + mobile breakpoints; page-worker agents produce a structured Section IR; deterministic codegen writes a real Next.js App Router project. Follow-up chat edits target a single page/section and regenerate only that slice.