Hi, I'm Aditya.
AI engineer working on LLM fine-tuning, quantization, on-device inference, and agent systems — and a Rust compiler contributor. I graduated from NIT Trichy in 2026, shipped production Rust + TypeScript/React as an AI developer on the 5-engineer founding team at Runitup, and I like building things that hold up in production, not just demos.
More about me · Reach me at adipras1407@gmail.com
-
AgentReplay — turns production AI-agent failures into verified pull requests. Ingests OpenTelemetry / LangSmith / Langfuse traces, auto-triages errors, loops, and unsafe operations, then reproduces failures in sealed Firecracker microVM sandboxes (zero production-credential exposure) and ships only regression-tested fixes. Supports LangGraph, OpenAI SDK, and CrewAI. In private beta.
-
Veredic — a security framework that applies formal verification to AI-agent protocols, combining runtime monitoring with static analysis to catch and prevent vulnerabilities in real time with minimal overhead.
Merged PRs in tooling used by millions of developers — full list on Notion:
- rust-lang/rust — fixed an ICE (compiler crash) in import suggestions, plus regression tests covering two further ICE classes.
- rust-lang/rust-analyzer — 4 merged PRs: configurable
dyninlay hints and three assist/diagnostic correctness fixes. - astral-sh/uv — extended pip compatibility (remaining file extensions,
--emit-index-urlde-duplication).
-
Sunny — a private, on-device skin tracker. Fine-tuned a multimodal LLM (SmolVLM-500M, LoRA/PEFT) on 2,292 clinical images, cut the footprint 93% (5.86 GB to 412 MB) via GGUF quantization, and ran it fully offline on Android with
llama.cpp— photos never leave the phone. 100% output-schema and safety compliance via grammar-constrained decoding (GBNF). -
Codesm — a multi-agent LLM coding system running up to 10 specialized subagents (coder, researcher, reviewer, planner, finder) across four model providers, with MCP client + LSP integration and a custom evaluation harness. Live at codesm.adityaps.work.
-
tryaudex — scoped, short-lived cloud credentials for AI agents. Wrap any command with temporary IAM permissions that auto-revoke on exit (AWS STS, GCP, Azure, Vault), HMAC-chained audit log, MCP integration, zero secrets on disk. Live at tryaudex.adityaps.work.
-
Glintly — describe an app in plain English and watch an AI agent build it in real time. Live at glintly.adityaps.work.
-
Piq — a local-first bucket-list app for Android. Dump things you want to try, shake your phone, do the thing. Flutter + SQLite, zero accounts, zero cloud. Live at piq.adityaps.work.
-
Visirs — a Rust library to visually group images and videos.
Languages: Rust · Python · TypeScript · Go · SQL AI/ML: PyTorch · LLM fine-tuning (LoRA, HF Transformers/PEFT) · GGUF quantization · llama.cpp · on-device inference · grammar-constrained decoding · agent orchestration · MCP · Weights & Biases Systems & Web: Axum · PostgreSQL · Next.js/React · Docker · Firecracker · AWS/GCP/Azure · CI/CD