I build systems that know what they can’t prove. And stop.

The discovery pipeline, in 3D

CareerOS reads six public job boards, scores every posting against a vault of 38 verified claims about what I have actually done, writes a tailored résumé for the ones that clear the bar — and then stops and waits for me. Nothing it produces is ever sent without my reading it first.

This page renders one real run as ten thousand particles. Each is a posting; each dies at whichever filter rejected it. The beam’s width is the survivor count on a log scale, so the narrowing you watch is the measurement rather than a shape I drew. Raw WebGL2, one draw call, no libraries.

Python  · FastAPI  · SQLite  · Next.js  · WebGL2

Open the pipeline →

Custody — proving what an AI did to a loan

Fannie Mae Lender Letter LL-2026-04 took effect on 6 August 2026. Any lender using AI in origination or servicing must now govern its use, extend that governance to its vendors, and — on Fannie Mae’s request — promptly disclose what AI it runs, for what purpose, and what safeguards are in place. You can answer that with a document describing what you intend to happen, or with a record of what did happen that the person asking can verify themselves.

Custody produces the second. The model call is the log entry, so the record cannot be forgotten; a deterministic gate acts as a live safeguard and records what it concluded; and the whole chain is hash-linked and Ed25519 signed, so a recipient can detect tampering without access to any of the lender’s systems. The demo runs a synthetic loan through five AI steps — including one where the model invents an income figure and gets caught — then lets you tamper with a record and watch your own browser find it.

Python  · SHA-256 chain  · Ed25519  · SQLite / Postgres  · WebCrypto  · pip installable

See Custody →

The containment gate

A language model writes the résumé bullets. This is what decides whether they ship. Every figure must appear in the source claim, every proper noun must be a tool or employer already on record, and verbs are treated as equivalence classes so that coordinated → led is free while applied → drove is not.

It has already earned its place: on a live draft the model invented a 35% figure that appears in no claim, and the gate discarded the whole generation. Anything it cannot judge routes to a human rather than guessing — a gap in the lexicon is the gate’s problem, never the writer’s. Type your own sentence into it and watch the checks run.

Claude API  · lexical + semantic verification  · deterministic scoring

Try the gate →

Mortgage GL reconciliation

A reconciliation specialist spent two to three hours every day cross-referencing three disconnected systems — Power BI, the loan origination system, and the A&B ledger — to classify and route GL discrepancies across more than a hundred open cases. None of it was written down anywhere.

Getting it out of one person’s head was the actual work: structured interviews and live training sessions, turned into engineering requirements, then into a production pipeline in Python, SQL and PySpark with statistical matching over thousands of transaction records.

The data-fetching layer underneath is deliberately LOS-agnostic, with automatic system detection, so the same automation runs identically across Encompass, MeridianLink and Blue Sage. That was not future-proofing for its own sake — an LOS migration was already in flight, and a rewrite mid-migration was the thing to avoid.

Python  · SQL  · PySpark  · Power BI  · Encompass  · MeridianLink  · Blue Sage

Ginnie Mae pool delivery

SettleDesk went to production for Ginnie Mae pool delivery and was live on its first deployment: 4 pools, 47 loans, $18.2M UPB, with 74 pending final certifications.

Underneath it, PoolPath merges ULDD XML loan exports from three different LOS platforms into valid Ginnie Mae PDD XML for SFPDM submission, correcting schema errors automatically instead of bouncing the file back to a person. A separate parser reads settlement and commitment dates out of unstructured warehouse-bank correspondence and populates the post-closing intake pipeline from it.

Claude API  · ULDD / PDD  · MISMO  · XML schema validation  · GNMA SFPDM

Enterprise CI/CD platform, security stage

I own the security stage of a five-stage CI/CD platform at a regulated mortgage lender — CodeQL, secret scanning and dependency review running in parallel, which halved pull-request check time from 18 minutes to 9.

The gates are built to prevent rather than report: secret scanning with push protection, a CRITICAL-severity CVE that fails the build, and required-reviewer branch protection. Azure infrastructure is Bicep — Container Apps, Key Vault, Container Registry, user-assigned managed identity — authenticated through OIDC federated credentials with no stored secrets, deploying incremental-only to protect brownfield resources.

The pipeline holds zero write access to IAM. Humans provision identities and secrets by hand, because a regulated lender will not allow self-service permissions, and the least-privilege model had to be designed around that constraint rather than argued with. Any engineering team onboards to the whole platform by copying three files and raising one pull request — same day, against an estimated two to three weeks.

GitHub Actions  · CodeQL  · Azure  · Bicep  · OIDC  · DevSecOps