Research and personal tools.

Alzheimer’s research data curation

Research in progress

Making public gene-expression data easier to audit and reuse.

I’m developing a tool to examine the data behind Alzheimer’s studies, preserving source evidence while flagging uncertain labels, repeated donor records and ambiguous gene annotations. A published analysis provides a case study for evaluating how verified curation decisions affect downstream results.

Work completed

• Acquired and inspected six public datasets containing 1,168 microarray measurements, with recorded sources, retrieval details and file hashes.
• Built checks for missing metadata, repeated donor and brain-region records, and incomplete gene mappings, while retaining original values.
• Completed an initial reconstruction of a published analysis, documenting its assumptions and an unresolved difference from the reported result.

Next steps

Validate the tool’s findings against reviewed source evidence and an existing curation workflow, then measure the consequences of supported curation decisions while holding the remaining analysis fixed.

• Python
• R / limma
• NCBI GEO
• Data provenance

BookGym

In personal use

Gym schedules and automated class bookings.

I built BookGym to turn recurring gym-class selections into scheduled bookings. A Python service uses Playwright to work with the gym’s existing website, while a calendar lets users manage their schedule. It runs on a Raspberry Pi for personal use.

Built so far

• Recurring weekly schedules, one-week skips, pauses, and seat preferences managed through a responsive calendar.
• Persistent booking jobs and event history distinguish confirmed reservations, waitlists, failures, and attempts that need review.
• Passwordless sign-in and delegated access, with changes attributed to both the acting user and the member being managed.

• Python / FastAPI
• Playwright
• SQLite
• Scheduling
• Raspberry Pi

Schema Studio

In-progress prototype

Healthcare source-to-canonical mapping.

An in-progress prototype for mapping healthcare source data into a canonical model. LLM-generated proposals go through explicit human review before deterministic execution.

Built so far

• Synthetic dataset: 12 CSV files and 24 canonical entities across Practitioner, Supplier, SupplierLocation, and PractitionerRole.
• Mapping proposals, human review, deterministic execution, provenance, and quarantine for records that need attention.
• A focused model explorer; live model integration has been tested.

Next steps

Continue testing proposal quality and review workflows against the synthetic dataset. The prototype is not a production integration.

• Source-to-canonical mapping
• Human review
• Deterministic execution
• Provenance

Fieldnotes

In-progress prototype

A local workspace for job-search evidence and drafts.

An in-progress local workspace for organizing a job search, connecting an evidence bank to drafting, and reviewing portfolio changes before using them.

Built so far

• A job-search queue and evidence bank.
• Review of portfolio content changes through a read-only importer.
• Versioned draft snapshots with rule-based drafting; drafting is not connected to a live LLM.

Next steps

Refine the evidence-to-draft workflow and review experience. Applications remain a manual step.

• Evidence bank
• Portfolio review
• Versioned drafts
• Local workspace