About

• Name: Alexandra Valentine-Ketchum
• Location: 🇬🇧
• Human Languages: 🇬🇧, 🇯🇵
• Programming Languages: Python, Rust
• Email: snowdrop404 🄰🅃 🄿🅁🄾🅃🄾🄽🄼🄰🄸🄻 🄲🄾🄼
• GitHub: snowdrop4

Self-taught programmer. I love safe, secure, fast, and robust code. I enjoy working on hard problems.

Technologies

• Data: XGBoost, scikit-learn, NumPy, SciPy, Pandas, Polars, PyTorch, Matplotlib, Seaborn
• Python Tools: Cython, Numba, Pydantic, Pytest, Mypy, Ruff, uv, Poetry, setuptools, functiontrace
• Rust Tools: Hotpath, PyO3, plotters, rstest
• Dev Tools: Git, Fish Shell, GitHub Actions, Ubuntu
• DevOps: Docker, Kubernetes, Helm, Argo CD
• Distributed systems: Prefect, Dask
• Observability: Grafana, Prometheus, Mimir, Alertmanager
• Databases: MySQL, Postgres, SQLAlchemy, Django ORM

Employment

fluidstack.io

• Period: 2025–2026
• Titles:
  • Forward Deployed Engineer (2025/01–2025/05)
  • Team Lighthouse Lead (2025/05–2025/11)
  • Software Engineer (2025/11–2026/05)

I worked directly with AI labs, maintaining and operating GPU clusters comprising tens of thousands of GPUs.

• Led the design and development of Kubernetes-based software to automatically detect and remediate hardware/software faults, then revalidate affected nodes. This consisted of continuous passive monitoring (observing hardware and operating system state), as well as periodic active checks (stressing InfiniBand and GPUs) across single and paired nodes. It also supported a mode for whole-cluster burn-in and validation for new deployments, and a CLI for manual diagnostic work. It was deployed in production across thousands of nodes in clusters owned by different suppliers, running on various different hardware, software, and network configurations.
• Led the observability team (Team Lighthouse).
  • Maintained Alertmanager, Grafana, Prometheus, and Mimir in production across tens of thousands of GPUs. Responsible for uptime and scaling whilst ingesting 1.2 million samples per second. Implemented various cost-reduction techniques (cardinality reduction, metric dropping, rewrite rules, container right-sizing, and so on).
  • Maintained alert definitions for the FDE team's on-call schedule (covering all aspects of the hardware and software stacks in GPU clusters), and the observability team's on-call schedule.
  • Maintained runbooks for every single alert in the FDE team's and observability team's on-call schedules.
  • Created a suite of cohesive and interlinked Grafana dashboards covering Kubernetes, Slurm, GPUs, CPUs, OS, Disk, Ethernet, and InfiniBand; with views aggregated by job, user, GPU, workload, container, and node, as part of a managed observability offering to customers.
  • Handled sales pitches to new customers for the managed observability product, as well as ongoing customer relations with requests.
  • Extended the open-source Slurm exporter with custom metrics to support a set of uniquely detailed Slurm dashboards.
• Worked closely with customers as a Forward Deployed Engineer, diagnosing and solving infrastructure issues during training and inference, across all kinds of operating system, hardware, networking, and software faults.
• Helped write a tool to manage datacenter buildouts: handling site scouting & acquisition, permitting, power projections, staffing, project management, inventory & BOMs.
  • Increased efficiency of LLM-aided development through various agent prompt files, spec files, agent hooks, CI-based agents for automatic PRs, and careful shaping of code to enforce business invariants at the type and test level.

chess.com

• Period: 2022–2024
• Titles:
  • Intern (2022/09–2022/10). Offered a full-time position halfway through the internship.
  • Core Chess Engineer (2022/10–2023/12)
  • Senior Engineer (2023/12–2024/11)

I worked in the cheat detection & research team, conducting research, rewriting legacy systems, and optimising detection software. Precise details are under NDA.

• Rewrote a cheat detection system from Perl to Python (wrote 80% of the code). Increased performance 4x. Discovered and fixed multiple long-standing bugs in the process. Created complete test coverage with over 1000 tests. Smoothly deployed as a drop-in replacement in production, maintaining bug-for-bug compatibility, and exactly replicating over 100 cheat detection statistics.
• Led development and design of a ground-up rewrite of a cheat detection research tool (wrote 80% of the code). Made it work as a distributed system with caching, to enable it to process hundreds of statistics per chess position across hundreds of millions of positions in minutes. Increased performance compared to the old system by 150x, even without the cache. Made it configurable and flexible enough to allow experimentation with different cheat detection methodologies.
  • After doing the rewrite, led and mentored a team of 4 on maintenance and further development. Wrote and managed tickets. Worked out requirements. Managed the project and its priorities, and set the short- and long-term direction.
• Researched and developed a machine learning model to predict how difficult chess positions are for humans. Researched and extracted thousands of features from a C++ chess analysis program, and then integrated the compiled model back into the C++ program. Built a program to empirically verify the model by testing it on chess.com employees.
• Collaborated with data scientists and chess experts on cheat detection. Implemented and optimised new and existing methodologies, identified bugs, proposed and researched new approaches, and discovered factors influencing detection power.
  • Invented a novel modification to a well-known statistical hypothesis test underlying the cheat detection system. Invented novel inputs to use in this hypothesis test, and used them both in combination to significantly improve cheat detection performance.
• Solved various problems with transpositions in a database of chess openings (forming a cyclical graph) that had vexed multiple people (including consultants brought on to tackle the job). Created a library to help with navigating and manipulating the data, and numerous scripts to fix the data.
• Served as the go-to person for helping with Python problems, and as a leader on Python best practices. Wrote and maintained the documentation and guides for working with Python at chess.com. Wrote and maintained the template repo for Python projects. Worked extensively on CI/CD and devex for various Python repos.
• Wrote a library for chess game analysis at scale. It supported using local chess engines, remote chess engines, and proprietary chess.com game analysis software. Made extensive use of async, websockets, multithreading, and multiprocessing for maximum performance. Used across various Python projects at chess.com.
• Ported academic code from a paper for the quantification of multimodal distributions from R to Python. Optimised with Numba. Increased performance 10x, and made it maintainable, sane, and verifiably correct with full test coverage.
• Created a chess engine compile optimiser that enumerates different compilers and different compiler flags, and empirically verifies the performance of every resulting binary. Identified builds that were 3–15% faster than those previously used.

Contracting

SkySight

• Period: 2024–2025

• Optimised, added tests, and updated meteorological forecasting code to modern standards.

Open Source

Ruff

• [pep8-naming] eliminate false positives for single-letter names (N811 & N814)

Reqsign

• Google Cloud: Add support for impersonated service accounts

OpenBench

• Ensure that workers don't re-use openings unless unavoidable

Personal Projects

Spooky-[go|chess|connect4]

Board game engines with Rust and Python bindings. Suitable for ML & DL applications.

• snowdrop4/spooky-chess (In use at chess.com!)
• snowdrop4/spooky-go
• snowdrop4/spooky-connect4

Tanmatsu

Declarative TUI library, bringing modern web layout features to the Unix terminal.

• snowdrop4/tanmatsu
• tanmatsu.readthedocs.io/en/latest/

Acorntorrent

Next-generation private BitTorrent Tracker.

• snowdrop4/acornserver

Achievements

• Ranked top 0.08% (~3.2 sd) of the competitive player-base in Europe in Overwatch (4200 SR).
• Ranked top 0.37% (~2.7 sd) of the competitive player-base in Europe in League of Legends (Diamond II) in my first season of play. Placed 4th on the op.gg Europe leaderboard for my character for that season.
• Ranked top 0.6% (~2.5 sd) of the competitive player-base in Europe in Deadlock (Ascendant 2) within a few months of play. Placed 163rd on the Europe leaderboard for my character.