computer engineering at ntu, singapore
rohan
kulshrestha
i design and build end-to-end products and ai solutions, and make music for fun.
selected work
all projectsteaching a small model to show its work: training an auditable research agent
A 9-billion-parameter open model that answers investment-research questions by looking things up in SEC filings through six tools, doing the maths explicitly, and citing the evidence for every number. Trained with imitation then reinforcement learning, and scored by code that checks every number against the filings.
mapping your skills to realistic career moves
A Singapore career guide that maps a person's skills to realistic next roles, skill gaps, live courses and matching jobs, using official SkillsFuture data with OpenAI as the interpreter and a 3D role sphere. Won 1st at the OpenAI x AI Singapore x PyCon Singapore 2026 hackathon.
shared context for coding agents
A plugin for Claude Code and Codex that gives every coding agent on a repository a shared room: they see each other's uncommitted edits, announce changes before making them, and preview the merged result before anyone commits. Placed third at the OpenAI x OpenRouter x ClickHouse x AI Tinkerers hackathon.
an agent that runs its own ml experiments
An LLM agent that improves a short-video recommendation model on its own: it reads the last training curve, picks a published technique, writes and runs the training code, and keeps the change only if it beats measured noise. Built by a team of four on the KuaiRand-Pure benchmark.