hi, i’m anish.

I study computer science at UC Irvine. I spent this summer as a software engineering intern at Roblox, and on the side I build LLM agents.

Hold, then let go at the top of the meter.

about

I’m a CS student at UC Irvine, graduating in 2028. I spent this past summer at Roblox, building the system that publishes a creator’s item to the catalog as soon as moderation approves it.

Most of my free time goes into LLM side projects. Right now that’s PaperMind, an agent that explains ML papers at whatever level you’re reading at.

When I’m not coding, I’m probably playing or watching basketball or tennis, or eating something with Nutella on it.

lately i’m into

  • Inference

    Serving models faster and cheaper, with KV-cache reuse and vLLM.

  • Post-training

    Fine-tuning open models with SFT and DPO.

  • Retrieval

    Hybrid search, and measuring where each method fails.

  • Evals

    Faithfulness, calibration, and LLM judges worth trusting.

  • Agents

    Tool use, MCP servers, and many agents arguing it out.

  • Distributed systems

    Workflows, retries, and idempotency.

experience

  1. 2026 Jun–Sep

    Roblox

    Software engineering intern, San Mateo

    Built auto-publish for UGC creators end to end: once moderation approves an item, it lands in the catalog on its own.

    Hover a step to see what it does.

    • Designed and deployed a Dockerized gRPC service for creator publish preferences and per-item opt-ins.
    • Built the Temporal workflow and C#/.NET publish activity, with UUIDv5 idempotency keys so Kafka retries never double-publish.
    • Built the React/TypeScript dashboard and REST API for publish defaults and fee consent, shipped behind a staged feature flag.
    • C#/.NET
    • gRPC
    • Temporal
    • Kafka
    • Docker
    • React
    • TypeScript

  2. 2025–26 Jun–Jun

    AI @ UCI

    Projects team

    Designed starter projects for club members to build.

  3. 2025 Jan–Apr

    ReadyTutor

    Software engineering intern, Irvine

    Built CourseEater, a UCI course scheduler with 500+ users, on a 10-person team.

    • Shipped 10+ full-stack features on MongoDB, Express, React, and Node.
    • Integrated the Anteater API for live course data, prerequisite checks, and grade distributions.
    • Cut schedule-building time 30% by tuning frontend rendering and backend queries.
    • MongoDB
    • Express
    • React
    • Redux
    • Node.js
    • Vite

projects

PaperMind reading Attention Is All You Need, with the outline on the left and a question panel on the right

PaperMind

An agent that reads ML papers with you and explains them at your level. Answers cite the exact lines they come from, and it points out the prerequisites you’re missing first.

  • Hybrid BM25 + dense retrieval hits 94.8% recall@10 on a 174-question benchmark over 8 papers.
  • Distilled Claude’s answers into Mistral 7B with QLoRA SFT, then DPO, served with vLLM. Calibration accuracy went up 18%.
  • A hand-written ReAct loop drives three MCP servers, including a Neo4j graph of 359 ML concepts.
  • Claude
  • FastAPI
  • FAISS
  • Neo4j
  • Redis
  • vLLM
  • React

Code on GitHub

A figure Lumi generated: the GLP-1R signaling cascade in neurons next to a heatmap of evidence strength for neuroprotection in Parkinson's disease

Lumi

A multi-agent virtual lab for drug discovery. Specialist agents investigate a research question, and an adversarial review panel scores how confident the findings are before they reach the report. The figure is one Lumi made on its own.

  • Built 4 MCP servers (BioRender, AntV, MockFlow, PyMOL) and 175+ tools across genomics, literature, and chemistry APIs.
  • Low-confidence findings go to human experts over Slack, backed by a SQLite world model that persists across sessions.
  • Claude
  • FastMCP
  • Pydantic
  • asyncio
  • SQLite
  • Next.js

Code on GitHub

say hi

Email is the fastest way to reach me.

anambira@uci.edu