Yash Budhia, Bengaluru

I build agents that have to keep running when nobody is watching. Always on, on real devices, with real user data behind them.

• Building Macy, an always on personal assistant. A Wear OS app in Kotlin does wake word detection and vitals on device. A Node/TypeScript server runs the agent itself: streaming voice, long term memory, a few hundred tool integrations. iMessage is the front end.
• The part of agent work I care about most is the boring part. Evals, MCP tooling, guardrails. Proving the agent picked the right tool, not just that it produced text.
• On numerical work I verify first and optimise after. moldcool checks every result against a series solution, a manufactured solution and an eigenvalue before it is allowed to report a speedup.
• Open source: 5 open PRs upstream (python-stdnum, Razorpay Python and Node, ibantools) and 2 merged into DiceDB.
• A lot of my recent work sits in private and client repos. What is below is the part I can show.

Agents and LLM systems

Backend and infrastructure

Apps and interfaces

Systems and numerics

A benchmark for code search agents. It builds the test set out of git history, so the files a commit actually touched become the answer key, and then scores any retrieval agent on weighted F1 and end to end latency. Agents are pluggable, there is a semantic search baseline to compare against, and LLM query rewriting is optional. Dataset generation, evaluation, metrics and reporting are separate modules, so plugging in a new agent means implementing one interface.

A Claude Code / Codex Agent Skill that teaches a coding agent to build gasless Solana onboarding without getting the sponsor wallet drained. It covers fee sponsorship, paying ATA rent for users, fee in token, and session keys. The main file is a threat model plus a checklist for the relayer: simulation, deny by default instruction allowlists, balance delta guards, compute caps, rate limiting, kill switch. Ships with a reference relayer that runs and has tests. Written for the Solana AI Kit skill bounty, after checking that nothing in the kit covered account abstraction.

A personal finance app that plays like a management game. The AI operator edits your records through typed, validated tools instead of guessing at free text. Bank statements (PDF, CSV or image) get parsed into a spending breakdown. Every change is logged and the last one can be undone. It degrades cleanly: a deterministic command parser when there are no API keys, an in process store when there is no database.

A 2D transient heat conduction solver that works out injection mould cooling time from first principles. The handbook Fourier formula is fine for a flat plate and wrong for ribs and thick sections, and this shows exactly where. Every number is checked against something exact: a series solution, a manufactured solution, the semi discrete eigenvalue, an energy identity, Eigen's direct solver. Convergence orders and stability limits are asserted in CI instead of eyeballed.

The results that were not flattering are in there too. OpenMP only gets 1.3x because the kernel is memory bound, and the shared memory tiled CUDA kernel is slower than the plain one on Ampere. Both kernels are kept.

An adaptive boss behaviour system for games. A RAG loop over FAISS and embeddings pulls context, the boss acts, and the outcome feeds back so it plays differently next time. Handles multiple games with their own vocabularies. FastAPI with Postgres and Redis, JWT auth, encrypted credential storage.

An image signal processing pipeline written end to end: demosaic, white balance, denoise, sharpen. Plus a deep learning denoiser and an LDR to HDR converter, with acutance and SNR numbers comparing the classical path against the learned one.