I'm co-founding Unizel. We build simulated companies where AI agents do real accounting work, graded to the cent.
Alongside that I publish small, finished tools built on one idea: let the model propose, let code check it, and send the unclear cases to a person. Computer Science and Economics, Case Western Reserve University, 2026.
| Project | What it does |
|---|---|
| billing-support-assistant (live demo) | Billing support where the AI proposes refunds and cancellations, code checks each one against the customer's records and billing rules, and unclear requests go to a person. |
| tripwire (results) | Adversarial test harness and independent deny-by-default action gate for tool-using AI agents: 44-class attack catalog, seed-reproducible runs, hash-chained audit log. |
| acquisition-launchpad | Phase 1 of a finance onboarding tool for acquired companies: human-reviewed account mapping, exact trial-balance tie-outs, readiness checks and a hash-chained audit trail. Sample data; later phases are planned, not built. |
| rescue-desk | Turns an ERP contract PDF into page-cited facts, has a person accept, correct or reject each one, calculates what is still owed with fixed formulas, and exports a review packet. Runs locally. |
| faultline (live page) | A small orders service instrumented with OpenTelemetry: flip a fault switch and the failing step shows up in the trace. |
| consumer-complaint-operations (live demo) | Dashboard of 84,194 public CFPB checking and savings complaints from 2025 that shows which patterns to review first. Python, SQLite, SQL, Chart.js, plus a Tableau version. |
Demos run on synthetic or public data, and each repository says which.
More projects
Agent evaluation and review (synthetic data)
- ask-once (demo): support inbox drafter that drafts replies from help pages and holds risky requests for a person.
- expert-test-round (demo): pre-launch expert test round for a product-support agent, where every miss gets an owner.
- shadow-run-scorecard (demo): decides when a new shipper's AI-entered orders can skip human review.
- benchmark-failure-readout (demo): splits the gap between two APIs by company type and failure reason.
Review queues and monitoring
- issue-triage (demo): review queue for public GitHub issues from n8n, Grafana and Kubernetes, with source links, evidence and CSV/JSON handoff.
- maintenance-checker (demo): checks CoreWeave's public status page and RSS feed for conflicting maintenance details and builds a review-gated calendar.
- kairos-watchers: credential-free GitHub Actions cron that polls 10 public feeds and commits a content-hash observation when one changes.
Go-to-market analysis
- gtm-signal-engine (demo): ranks accounts by buying signal from public ATS job boards, with no API keys and no paid data vendor. Snapshot from August 8, 2026.
- reseller-payout-checker (demo): reseller rev-share payout checker with whole-cent math and held exceptions (synthetic data).
- signup-account-rollup (demo): rolls developer-API signups up to companies and picks the accounts worth a conversation (synthetic data).
- trades-hiring-signals (demo): ranks home-service shops to call first from job posts (synthetic data).
- channel-adoption-readout (demo): which pages bring signups who keep running workflows (synthetic data).
- channel-cac-ledger (demo): cost per paid practice by channel, turned into next month's budget (synthetic data).
- lead-feedback-readout (demo): readout of the leads sales reps turn down (synthetic data).