Resume

Alex Ashworth.

LLM product engineer.

Shipped an AI travel product end to end as its only engineer, from research agent to billing. Before that, built evaluation tooling and data pipelines for an ask-your-data analytics product. Now contributing full time to open-source Rust and looking for an engineering role.

Belmont, NCAI products / Open-source software

Experience

Building, shipping, improving.

clonk-rs

Current

Contributor/Open Source

Jul 2026—Present

Rust contributor to clonk-rs, an open-source rebuild of a 2D destructible-world game engine: 172 pull requests merged in eight weeks across 1,588 files.

Cut lobby join time from 9.0s to 0.47s by admitting hosts before exact resource packing, and 20ms/frame by caching the sky instead of relighting it per frame. Removed quadratic object resolution from mouse picking.

Work split 114 bug fixes, 21 performance changes, 14 test changes across the renderer, network layer, and script compiler.

172PRs merged in eight weeks
9.0s → 0.47sLobby join time
  • Rust
  • Rendering
  • Networking
  • Compilers

Tripiage

Technical Founder/Full-time · Remote

Sep 2025—Mar 2026

Sole engineer on an AI travel-planning SaaS. Split the system into a Next.js application and a research worker so multi-minute jobs ran off the request path.

Drove research with a ReAct loop under a per-job token budget, and enforced citations in code so uncited claims never reached an itinerary. BullMQ and Redis queues, PostgreSQL.

Launched publicly with paying subscribers on Stripe billing; the company wound down in March 2026. 88 tests, weighted toward authentication and rate limiting.

  • Next.js
  • LLM agents
  • PostgreSQL
  • Redis
  • Stripe

Genetica

Machine Learning Engineer/Full-time · Remote

Oct 2024—May 2025

Built an ask-your-data chat interface over customer databases with one other engineer. Implemented M-Schema to give the model an accurate view of database layout.

Built the evaluation harness: commenting !eval on a pull request benchmarked its prompt or model against the deployed baseline for accuracy against a direct-SQL reference, consistency, and instruction following.

Replaced 1-2 hours of manual spreadsheet enrichment per dataset with a 10-second asynchronous pipeline, no human in the loop. Cut Snowflake spend over 50% through auto-suspend tuning and warehouse right-sizing. Sustained a 10,000-user load test.

  • LLM evaluation
  • Data pipelines
  • SQL
  • Snowflake

Projects

An idea, taken all the way.

Blujay

Searchless Transformer Chess Engine

Play Blujay

Sole Developer·2025 — Present

Trained a searchless transformer chess engine that predicts action-values directly, after DeepMind's design. 8 layers, model dimension 256, 128-bin value head, trained 23 epochs over 13.5M positions in roughly 35 GPU-hours.

Reached 19.8% exact-bin validation accuracy against a 0.8% chance baseline; playing strength remains weak, so the model is a research prototype.

Built the data pipeline: 15 parallel Stockfish workers under health checks and retries, filling gaps in 50M Lichess evaluations. Lazy loading cut peak memory by 70%.

  • PyTorch
  • Transformers
  • Python
  • Chess

Education

University of Minnesota: Twin Cities

B.S. in Computer Science

Sep 2020—May 2024

Additional coursework

fast.ai, Practical Deep Learning for Coders

Dec 2022 - Jan 2023

Skills

The tools behind the work.
Languages
  • Python
  • TypeScript
  • Rust
  • SQL
  • Java
  • C/C++
Frameworks
  • PyTorch
  • FastAPI
  • Next.js
  • React
  • Node.js
  • Pandas
  • Transformers
Infrastructure
  • Docker
  • Kubernetes
  • PostgreSQL
  • Redis
  • Kafka
  • Snowflake
  • AWS

Beyond work

Japanese

Four semesters of college Japanese. Can hold a basic conversation and read simple business Japanese.

Away from the editor

  • Kaggle competitions2022-2024
  • USCF-rated chess player, 1800+2021-present

Thanks for taking a look.

Let's talk