Please review the projects.

Metadata for a public dataset is scattered across the entry, the paper, and the supplements. A curator reads all three.

First author. I own the agent architecture, the field extraction, and the recall measurement.

• An orchestrator delegates retrieval, parsing, ontology mapping, and inference to expert sub-agents.
• Recall is measured across all 23 fields, including original terms, normalized terms, and ontology identifiers.
• The system must scale to hundreds of fields without losing precision.

• Preprint on bioRxiv, DOI 10.1101/2025.06.10.658658.
• 93% average recall across 23 fields, reported in the abstract.

Retrieval pipelines are only as good as the text and structure they ingest.

I own parser design, extraction boundaries, and the JSON shape.

• Extract sections, tables, figures, and references into predictable JSON.
• Keep the output useful for RAG pipelines and training data.
• Keep the source public.

• 98% extraction accuracy measured across more than 7,500 papers.
• Purpose-built for PubMed Central full-text extraction.
• Open source. Install it, point it at a paper, and check the JSON yourself.

AI engineering interviews cover math, ML, LLMs, RAG, agents, evaluation, LLMOps, safety, and system design.

I own the product shape, question bank, study loop, review behavior, and the landing page.

• Broad topic coverage must still feel navigable.
• Practice must force recall before the answer appears.
• Weak questions must come back without extra dashboard work.

• Public metadata lists 2,049 expert answers across 28 AI-engineering domains.
• One share brought roughly 650 organic users in the first three weeks.

Every cancer dataset carries its own field names. Pooling two of them is a translation job before it is a research job.

Contributor. I released 23,650 cancer patient phenopackets with NCI and Lawrence Berkeley Lab.

• The output must satisfy the GA4GH phenopacket schema, not a local convention.
• The conversion has to survive the gaps that real clinical records carry.

• Public repository under the Monarch Initiative.
• 23,650 released phenopackets, also listed in the public record on the about page.

The people a health scheme is written for are often the people least able to read the form that claims it.

I own the voice interface, language coverage, eligibility logic, and the false-positive budget.

• A caller needs an answer by phone, without an app and without literacy in English.
• The system covers 23 languages.
• A wrong yes is worse than a slow no, so false positives stay near zero.

• Public repository on GitHub.
• Claims here cover product shape and role only.

A flat-rate coding plan hides what it does. The useful question is which project, pull request, model, and session produced the work.

I own log parsing, rate-card pricing, evidence labels, the privacy model, and the CLI.

• Usage stays local unless the user shares a report.
• Cost numbers need dated rate cards and visible assumptions.
• The dashboard separates exact, estimated, partial, and unsupported evidence.

• README documents local-only operation across Codex CLI and Claude Code logs.
• Caliper installs through the caliper-ai package, with a reproducible demo command.

WHOOP charts show data. Training decisions need computed context: sleep debt, strain trends, load, and correlations.

I own MCP tool design, OAuth setup, the WHOOP API client, analytics, and docs.

• Health data stays read-only and local by default.
• The server computes trends before the model explains them.
• Setup makes sure that auth, scopes, client config, and API access work.

• README documents 24 WHOOP tools across recovery, sleep, strain, workouts, and trends.
• CI-backed Python package with offline tests against a faked WHOOP API.