Welcome 🗿

My general vibe:

•AI power user. Zellij, Pi, Claude, Zed, and too many agent windows open at once.
•Into AI infra, full-stack stuff, and making systems robust enough to just work.
•Quite studious and active.

View resume Connect on LinkedIn See the code

Lately

Work

AI engineer at Avathon. I own the agent platform behind the conversational AI product, the knowledge graph it reads, and the evaluation that keeps its answers correct.

UT Austin OMSAI

B.S. in CS & Math from UT Austin, now doing the online MS in AI part-time while working full-time. Currently taking AI in Astrophysics.

Gym, tennis & cycling

Usually try to be active 6× a week between lifting, tennis lessons, and cycling. Sometimes I go 6/6, other times I'm at 1/6. I play it by ear.

TV & reading

Recently finished Snowfall and Succession — currently watching Shrinking. Book-wise, just read Stoner and How Machines Learn. Always something in rotation.

Work experience

AI Engineer

Avathon · Oct 2025 – Present

⌄

I build the harness agents run inside: the tools they call, the enterprise knowledge graph they read, and the evaluation that keeps them correct — wired together into a multi-tenant platform running in production.

Built the knowledge-graph agent behind the Barrick Gold engagement, working directly with client leadership to shape what they valued. It anchored a multi-year enterprise contract now in production.
Extracted the company's conversational AI from a product monolith into a standalone multi-tenant agent platform (Django async ASGI, LangGraph, LiteLLM gateway) and shipped it through dev, UAT and production for a Fortune-500 deployment. It's now the company's shared agent layer.
Replaced config-trusted persona scoping with authorization enforced against live identity (Keycloak/OIDC), clearing the last blocker to general availability. Added drop-in MCP server support with per-server isolation and forwarded user tokens, so new tool servers go live by configuration with no code and no deploy.
Designed and materialized the canonical ontology behind the agent's answers, modeling 28.7M rows of client data into an agent-queryable knowledge graph, plus the operations layer that verifies freshness and correctness and runs dependency-ordered recompute.
Own agent quality with no dedicated QA function on the team: built the evaluation and observability stack (Langfuse, per-query cost/latency/token telemetry) and run continuous evals scoring answer correctness and tool-call trajectory across 16 identity-gated personas.
Fine-tuned Qwen2.5-14B (QLoRA) for natural-language-to-graph-query translation on 1,232 execution-verified pairs, reaching parity with Claude Sonnet on single-query generation at a fraction of per-query API cost.
Built the carrier scoring and tendering engine for the supply-chain product, separating metric computation from decision policy. Drove adoption of DBOS over Celery+Redis for crash recovery and audit trails.
Rebuilt and shipped avathon.com in two weeks (Next.js, Sanity CMS) with the VPs of Strategic Sales and Marketing and a peer engineer.

Data Science Intern

IBM · May 2023 – Aug 2023

⌄

Built and deployed a Flask proof of concept integrating Watson Assistant with Db2, used in the pitch that won the client engagement.
Shipped a Dash application on watsonx and Watson Discovery with a custom evaluation API for side-by-side LLM comparison.
Owned backend API integration for a React and Flask onboarding MVP on IBM Cloud and OpenShift.

Undergraduate Researcher — Earthquake Modeling

UT Austin · Jan 2024 – May 2024

⌄

Re-engineered a legacy Mathematica seismic-response solver into a multithreaded Python pipeline.
Processed 29k+ NGA West2 records and ran 64-vCPU GCP sweeps, cutting runtime from roughly 3 years to about 2 weeks.
Produced interactive Plotly dashboards and a seminar deck for faculty presentation.