Software engineer, AI/ML · UT Austin

Hi, I’m Varun.

I study computer science at UT Austin. I co-founded DUIBL, where I built Stilez, an app that rates your outfit with AI, and I interned at Cox Automotive in 2025.

Experience

Newest first. Each one links to a page with more detail.

  1. – now

    Co-founder & lead engineer Stilez Closed source

    DUIBL

    I co-founded DUIBL and built Stilez, an iOS app that rates your outfit with AI. I took it from the first architecture sketch to the App Store.

    • −67%image-gen spend
    • 2×rating throughput
    • 68%cache hit rate

    React Native · TypeScript · Postgres · pgvector · Deno · LLM pipelines

  2. –

    Software engineer intern vAuto Proprietary

    Cox Automotive

    A summer on vAuto, software used by 7,500+ dealerships. I worked on vehicle sourcing, deploy tooling and an internal auditing app.

    • −90%load time
    • −80%blue/green shift time
    • 1,000+inconsistencies surfaced

    React · AWS · Slack API · Blue/green deployments · REST APIs

  3. 2024

    Earlier

    A summer in UT Dallas's VR lab building a wind-tunnel simulation in Unity, and co-founding Growth Prep Academy, an AI test-prep platform that reached 500 beta users.

Projects

Side projects on LLM reliability, MLOps and distillation. Each one is measured against a baseline.

  1. –

    LLM reliability harnesses Open source

    Aegis

    Reliability harnesses for LLM apps, each tested under a protocol fixed in advance. In the latest study, a 140 ms decision model beat an LLM judge at catching prompt injection, at 1/18 the cost.

    • 0.994injection AUROC (judge: 0.980)
    • 4×faster than the LLM judge
    • 1/18the cost per check

    Python · MLflow · tau2-bench · Jev

  2. –

    Self-healing MLOps platform Open source

    Demand Forecasting Platform

    Forecasts retail demand, notices when its accuracy drifts, and retrains and redeploys itself.

    • −46%forecast error (WMAPE)
    • +71%accuracy after a demand shock
    • 206automated tests

    Python · LightGBM · MLflow · DVC · FastAPI · Docker

  3. Agentic computer-vision pipeline Open source

    DermAssist

    A LangGraph pipeline that takes a skin-lesion photo through preprocessing, a ViT classifier, Claude interpretation, and PubMed literature grounding, and never finalizes a report without human review.

    • 8pipeline stages
    • 7lesion classes (HAM10000)
    • 0reports without human review

    Python · LangGraph · ViT · Claude API · ChromaDB · MCP · Streamlit

  4. –

    Research Open source

    Distillation Reasoning Transfer

    Can a small local model learn to reason from a frontier model using only its outputs? Trained on a consumer GPU.

    • 30→46%GSM8K accuracy
    • ~46%of the gap to the teacher closed
    • 3,000filtered reasoning traces

    Python · PyTorch · Hugging Face · QLoRA · PEFT

  5. Early 2026

    Multi-agent research system Open source

    Stock Advisor

    A market-research system on Google ADK: an orchestrator delegates to data, technical, fundamental, and news-sentiment agents, then synthesizes a weighted recommendation in a streaming Streamlit UI.

    • 5cooperating agents
    • 10+technical indicators
    • 4analysis lenses, weighted

    Python · Google ADK · Gemini · pandas-ta · Streamlit

GitHub activity

Toolkit

Languages
Python, Java, C++, C, C#, Go, JavaScript, TypeScript, SQL, HTML/CSS
AI / ML
PyTorch, TensorFlow, Scikit-learn, Hugging Face Transformers, OpenCV, NumPy, Pandas
Agents & LLMs
LangChain, LangGraph, Google ADK, RAG pipelines, multi-agent systems, vector databases (pgvector, Chroma, Qdrant), fine-tuning, distillation, evaluation harnesses
Infra & DevOps
AWS (S3, Lambda, RDS, ECS), Docker, Git, CI/CD, GitHub Actions, REST/Express APIs, React/Next.js, PostgreSQL
Education
The University of Texas at Austin, B.S. Computer Science.
Data Structures & Algorithms · Operating Systems · Computer Architecture · Machine Learning · Artificial Intelligence · Natural Language Processing · Computer Vision · Generative Visual Computing

Say hello.

If you want to talk about any of this, email me.

varunvsaran@utexas.edu