Six years making data systems reliable. Now making AI systems auditable.
Before this I built quantitative dashboards at Fidelity and, earlier, ML pipelines and computer-vision MVPs. Now I'm doing AI safety research through SPAR on how outside auditors can check a lab's models, and building safety tools in public.
- researchSPAR: evaluating AI auditors, with Auditor-in-a-Box
- lookingfor safety- and alignment-focused roles, remote
Selected work
Auditing AI from the inside out
SPAR · in progressResearch in the Supervised Program for Alignment Research. If an insider quietly trained a model to serve a hidden interest, could an outside auditor tell from the lab's records alone? I'm building evaluations to measure that, run through Auditor-in-a-Box, where the auditor works inside a sealed environment and only an agreed verdict leaves.
Ground Knowledge
solo · liveA tool for contested research questions that weighs independent evidence instead of raw citation counts. It surfaces false consensus and source overlap, so twenty papers re-analysing one dataset count as one root, not twenty.
Experience
- Building evaluations of how well AI auditors catch a model that was covertly trained to serve a hidden interest, working from a lab's records alone.
- Running them through Auditor-in-a-Box, a protocol for third-party audits inside a sealed environment, to compare auditing in the box with direct model runs.
- Rebuilt a legacy quantitative dashboard portal (Python, Angular, AWS) for 200+ users and 50+ concurrent, replacing a system with silent data-pull failures and no version control.
- Re-architected deployment to decouple dashboards from the core portal, allowing R, Python, and Angular dashboards to scale without adding load to the portal.
- Cut page load times by up to 80% and raised automated test coverage to 90%; built a shared Angular component library and migrated legacy R apps to Python + Angular.
- Built an MVP video classifier end to end with FastAPI, Streamlit, and Docker.
- Engineered Python applications for portfolio optimization, turning heavy quantitative algorithms into maintainable code and cutting Monte Carlo runtime by 70%.
- Designed ETL pipelines for large-scale data, improving accuracy and reliability downstream.
- Ran impact analyses for global partners that shaped educational strategy for Minerva Forum users.
- Built an internal Streamlit analytics app that cut report creation time by 50% and supported 10+ configurable visualization types; mentored interns in Python.
- Taught 30+ students Python, ML, and statistics with a full curriculum, holding a 5.0 rating.
- As a Python data analyst at LucidMove, improved data-retrieval efficiency by 90% with optimized web-scraping and built predictive models on complex datasets.
Toolkit & background
- ai safety
- LLM evalsInspectthird-party auditing
- languages
- PythonJavaScriptRCSQL
- ml & data
- NumPyPandasscikit-learnMatplotlibSeabornPlotlyStreamlit
- backend / web
- FastAPISQLAlchemyAngularVue.js
- infra
- DockerAWSAzureTerraformGitJenkinsOracle
- methods
- ETL designstatistical modelingMonte Carloperformance optimizationpredictive modeling
Writing
Open to safety- and alignment-focused roles, remote-friendly.
If you're working on evaluations, auditing, robustness, interpretability, or governance tooling and could use an engineer who ships, I'd like to hear from you.