Hire me
Where I have worked and what I shipped there. Every number below is reconciled against my resume; where a figure is private, this page says so rather than estimating.
Tools, and where they were used
Derived from the eighteen projects in the catalogue rather than typed out here, so nothing on this list is a tool I have not shipped something with. What sits beside each row is the evidence for it, and a short row is a weaker claim than a long one.
- TypeScriptCocoon · Verso · unlinkd +6
- PythonFraudStream · Talker · AI-Powered Trading Bot +5
- Next.jsVerso · Tidy Roll · HealthCalc +4
- TailwindHealthCalc · Numen · Link Flame
- BunFeedless · HealthCalc
- FastAPIqsensor-sim · Album Conceptualizer
- PrismaRemedi · Link Flame
- ViteCocoon · unlinkd
- Alpaca Trading APIAI-Powered Trading Bot
- Apache KafkaFraudStream
- Celeryqsensor-sim
- Docker Composepi-lab
- ExpoTidy Roll
- FAISSTalker
- FinBERTAI-Powered Trading Bot
- GitHub APImerge-gate
- IndexedDBunlinkd
- LangChainBlog-AI
- Manifest V3Cocoon
- MDXBlog-AI
- Nginx Proxy Managerpi-lab
- NLPPaper Summarizer
- OLlama2Talker
- OpenAI GPT-4Blog-AI
- PandasAI-Powered Trading Bot
- Pi-holepi-lab
- PlaywrightFeedless
- PostgresRemedi
- pytestmerge-gate
- RAGTalker
- ReactCocoon
- React NativeTidy Roll
- SciPyqsensor-sim
- SnowflakeFraudStream
- Spark StreamingFraudStream
- SQLiteVerso
- TransformersPaper Summarizer
- VitestFeedless
- Web Cryptounlinkd
- WireGuardpi-lab
- ZodVerso
Work experience
Sizzle
Sizzle
- Deployed the pipeline on GCP Vertex AI and BigQuery, serving 5k+ daily inference requests in production.
- Integrated the fine-tuned vision model into a production REST API in Kotlin, holding a <200ms p95 target for 10+ B2B clients.
- Turned vague quality complaints into concrete labeling guidelines across a 50k+ image dataset, raising model accuracy by 12 percentage points.
Upwork
Upwork
- Delivered custom ML for 4 clients: RAG document Q&A, fine-tuned classification, and automated data pipelines.
- Built the MLOps around them — CI/CD, model versioning, monitoring.
- Maintained a 98% Job Success Score by setting expectations up front, demoing work in progress, and documenting the handoff.
G&M Trailer Repair, Inc.
G&M Trailer Repair, Inc.
- Rebuilt receipts, pick tickets, and job scheduling into searchable Excel systems and a Kanban board inside Microsoft Teams — the suite the crew already used.
- Built and owned the parts inventory, an hour log normalizing labor against per-client rates, and PTO accrual formulas, automating the repetitive steps in VBA and invoicing in QuickBooks.
- Coordinated repair intake and dispatch for a fleet running roughly 200 trailers a month, working daily with a primarily Spanish-speaking crew.
Joint Genome Institute
Joint Genome Institute
- Cut processing time by 90%, shortening iteration cycles for the research teams.
- Engineered ETL pipelines for genome reference data, cutting preprocessing from 4 hours to 10 minutes.
- Delivered a KNN classification model for plant-microbe interaction prediction (F1 0.68), and walked researchers through what that number did and did not support.
- Presented pipeline capabilities and architecture to 1,000+ end users and senior leadership at the annual user meeting.
VICE Lab
VICE Lab
- Parallelized model execution on Slurm HPC, cutting total runtime by 50%.
- Ported core Python logic to Cython for a further 30% runtime reduction.
- Built a Vue.js visualization layer stakeholders could read without an engineer's help.
How I work
Most of this field is locked behind academic papers and tooling that assumes you already know the answer. The work I like best takes something that only runs in a notebook and turns it into something a person can open.
Start simple, measure, then change it based on what the measurement said. That order matters more than which framework is involved.
If you can't explain it simply, you don't understand it well enough. Feynman said it first and I have not found a counterexample.
Current focus
RAG systems
Exploring different use cases for retrieval-augmented generation and making them more marketable. Basically trying to get paid for doing cool stuff.
Open source
Contributing to open-source projects and building tools that make data science more accessible. FOSS is how the internet should work.
The resume
The same history as a PDF, kept in sync with this page. A call works too if you would rather ask than read.
Hiring for a project? See how I work with clients.