Current work

Build close enough to the problem to change your mind.

Since leaving PowerSchool, I’ve deliberately gone deeper into direct building. The work ranges from product prototypes to AI systems, engineering experiments, and independent research. The common thread is staying close enough to the customer, system, and evidence that reality can change the plan.

Product builds

Start with the customer job.

These are the most direct extensions of the Product work: build enough of the workflow to test the premise, let discovery change it, and keep consequential output tied to evidence.

Working prototype · attorney discovery

Matter — legal-document workbench

A practicing attorney killed the initial intake premise. I rebuilt around the actual discovery/document job: ingestion, OCR fallback, source quality, conflicts, provenance, evidence gates, and a Clio-ready integration plan.

Read the case study →
Working software · pilot proposed

Spanish for Connection

A LiveKit/Next.js speaking-practice product for learners who understand Spanish but freeze in live conversation. I sourced a school directly, met with a board director and Spanish staff, and received verbal agreement to a free pilot in the first stakeholder meeting.

See the product story →
AI systems & evaluation

Build the layer around the model.

Handoff, provenance, evaluation, and human judgment: systems for keeping powerful generators useful without letting fluency silently become evidence.

Public release candidate

OXBOW — carry the work, not the model

A self-orienting handoff object that can orient a fresh model, carry a corpus and working context, verify what arrived, expose tools, and record new work without claiming continuity of the model itself.

77-item reference instrument

The Catch

A human-in-the-loop judgment instrument for distinguishing actual model-output failures from clean-but-suspicious output, with partly-right cases, provenance and solvability classes, signal-detection scoring, and referent-vs-echo probes.

Read the method →

AI systems & evals →

Engineering

The check ships with the claim.

I build with AI-assisted implementation and own architecture, decomposition, test design, failure boundaries, and release judgment. The artifacts below make that ownership inspectable.

Start here

Engineering Verification

Claims, checks, caveats, reproduction commands, and the places where the mechanisms lose. The verification record is the shortest path to the technical evidence.

Read the record →
Human perception proposes

Residual Lens

A self-contained browser instrument for seeing candidate structure across up to six scalar dimensions. Visible patterns get proposal rights, not proof authority.

All engineering work →

Research

Proofs, measurements, conjectures, and killed hypotheses stay separate.

Independent mathematical work is the deepest layer of the site. The same operating rule applies here: claim only the status the evidence supports.

Interactive companion

One sine, two numbers

A draggable companion to the shifted KMS work: the target triangle, affine singularity web, spectrum, and five-regime conditioning law computed live in the browser.

Explore the companion →
Structured matrices + prime number theory

Independent research

The research index keeps proved, empirical, conjectural, and killed claims separate and exposes the work at the depth the evidence supports.

Research index →