Section — Work5 live systems4 open casesSome held back
Products I still run, the engagements that shipped, and a shelf I can’t show in public.
p. 1
Each one exists because something annoyed me enough to build the fix.
Private writing companion — AI crafts prompts; you write the answers.
Why
Most “AI journals” write for you. That kills the point of a journal. I wanted a tool that only nudges — prompts shaped to your themes — while the words stay unmistakably yours.
What shipped
Theme-based prompt generation, private accounts, timed and spontaneous nudges. The model never authors the entry.
Multiple models deliberate, judge, and consolidate hard questions.
Why
One model on a hard question is a single overconfident voice. I wanted deliberation as a product — propose, judge, consolidate — so disagreement becomes signal instead of noise.
What shipped
Multi-provider completion, judging rounds, and a consolidation step into one answer you can actually inspect.
Experiment with prompts, compare models, iterate in one place.
Why
Prompt work usually lives in chat tabs and half-broken scripts. Tiny wording changes matter, and comparing models side-by-side shouldn’t require a new ritual every time.
What shipped
Edit prompts and params, run across models, compare outputs in one loop so iteration stays visible.
Review research through persona-aware lenses for HCI and GenAI work.
Why
HCI and persona papers need reviewers who understand representation risk — not generic “be nicer” feedback. I built PPR to stress manuscripts the way our research actually critiques them.
What shipped
Agentic PDF review, revision validation against prior comments, grep-grounded style checks without a page cap.
Playground for modeling persona ecosystems — minority views to whole populations.
Why
Personas from scattered research are slow, inconsistent, and easy to stereotype. PEP exists to turn unstructured context into structured persona ecosystems you can probe — including minority views.
What shipped
Ingest research/transcripts, retrieve with vectors, generate and expand persona sets grounded in your own material.
p. 2
Described by outcome, not by tool list.
Document verification and audit trails for a regulated fintech, built for GDPR scrutiny.
Credit risk models with explainability for Caribbean banking constraints.
Observability and log intelligence — anomalies, bursts, and drift for AIOps demos that behave like production.
Cost-aware, fairer resume screening by orchestrating multiple LLMs under uncertainty.
p. —
Client-sensitive. Checked on the server, not in this page.
A few engagements stay behind a gate — client-sensitive detail, not marketing copy. The passphrase is checked on the server; nothing confidential is in this page until it passes.