31 systems with something you can inspect

Stops a coding agent from leaking secrets, deleting data, or claiming success without proof — 79 tests, zero dependencies

A portable safety fence for AI coding agents. It blocks the dangerous actions — leaking secrets, deleting things, claiming done without proof, rewriting history — and lets the safe ones through. Off by default and on purpose.

Python · deterministic rules · fail-closed · 79 tests

Version-control record: 2026-05-29 → 2026-08-20 · 5 commits

A weak model or agent run leaves evidence and can stop the release instead of disappearing inside an average

A runnable model canary with structured performance telemetry, automated output checks, evaluator calibration, agent tool and loop controls, and one exportable evidence chain.

Next.js · structured telemetry · evaluator calibration · agent traces · SHA-256 evidence chain

You can see a long generation run going wrong while it is still cheap to stop

A status board that makes a long generation run auditable while it is still running, instead of after it has finished going wrong.

spec-driven pipeline · status board

A reviewer's decision becomes data the pipeline obeys — not a comment somebody may read later

Human-in-the-loop review as a first-class component: the reviewer's decisions are structured output the pipeline consumes, not a comment thread someone has to interpret later.

typed JSON · review gates · human-in-the-loop

3,400 exam questions crossed a language without changing meaning — and any batch that cannot prove it stays behind

Translation at volume where the interesting problem is not the translation but knowing which batches are safe to ship. Every stage is a gate with a written pass condition.

multi-model · staged QA gates · resumable runs

A couple and their therapist see the same week, and the couple decides what the therapist sees

A shared practice workspace for couples counselling. The demo runs both lenses: switch between the couple view and the therapist view and watch a sharing scope turned off disappear from the therapist's session brief on the next render. A couple who has not finished consent shows a consent boundary instead of a brief. Grew out of two precursor builds, The Palms and gotta-guy.

Next.js · consent scopes · role-scoped views · synthetic data

The learner has to explain their answer before moving on — which is where the actual learning happens

A Socratic tutor for the CSLB contractor licence exam: the learner picks an answer and then has to explain why, which is where the actual teaching happens.

Groq · Socratic prompting · exam prep

A food vendor sees which menu item is losing money while there is still time to change it

Food-vendor cost and margin calculator: edit menu items inline and watch per-SKU margin move. Unit economics a vendor can actually operate, not a spreadsheet they abandon.

Next.js · unit economics · margin modelling

The owner, the manager and the tech each see the one screen they need — and a slipping job surfaces before the customer calls

Role-aware field-service operations: portfolio KPIs and threshold alerts for the owner, a one-click assignment queue for the manager, today's jobs with photo-evidence capture for the tech.

Next.js · field ops · role-based UI · dashboards

One login gives the owner the briefing, the estimate and the quote — instead of four tools and a lost afternoon

Gated owner toolset: AI business briefing, walk-site estimator, quote builder, and admin controls in one place.

Next.js · Groq · estimator · quote builder

A book gets written in one workspace instead of three tabs — progress visible, drafts assisted, reading alongside

A themed chapter reader alongside manuscript tracking and an LLM drafting assistant, in one workspace rather than three tabs.

Next.js · manuscript tracking · LLM assist

After an incident, you can replay exactly where everything was and when — instead of reconstructing it from memory

Real-time asset tracking on an SVG floor plan, with breach alerts and a scrubber to replay what happened and when. Built for the moment after the incident, when someone asks where things actually were.

SVG mapping · geofencing · realtime · history replay

Every voter sees exactly why a candidate ranked where they did — and the same address always gets the same answer

Address in, ranked candidate matches out, drawn from 25+ public-record sources. Same inputs produce the same ranking every run — for anything touching an election, an unexplainable ranking is not a feature.

Next.js · TypeScript · determinism audit · 25+ record sources

Version-control record: 2023-07-04 → 2026-08-20 · 1,391 commits

Flashcards become a playable game only when the material actually suits one — it scores the fit instead of forcing all three modes

Turns flashcards into playable Match, Order or MCQ mini-games, and scores which mode actually suits the material rather than generating all three and hoping.

deterministic generation · suitability scoring · HITL playtest

Both partners see the same exercise at the same moment — and the guidance comes from published sources, not the model's imagination

Guided communication exercises with both screens synced live. The content is source-grounded rather than model-improvised, which is the whole point in this domain.

Supabase Realtime · Next.js · TypeScript

Any image it made can be made again — the settings are recorded, so a result is reproducible instead of lucky

Image generation as a repeatable pipeline with recorded settings, so a result can be reproduced rather than re-improvised.

image generation · workflow capture

Change the questions without rebuilding the form — and every submission arrives structured, ready to use

A JS object defines the sections, field types and conditional logic; the form and its typed output follow from it. Change the schema, not the form.

schema-driven · conditional fields · typed JSON export

Mismatched money surfaces where someone will actually see it — exceptions raised, not buried in a spreadsheet

Transactions matched and exceptions raised where someone will see them, which is the only place a reconciliation tool earns its keep.

TypeScript · reconciliation · exception surfacing

No model grades its own homework — three propose, plain rules rank them, and a person makes the final call

One item routed to Ollama, Groq and Codex at once. A deterministic scoring layer picks the winner, then a person approves before anything is written to SQL. The cost work is the governance work — controls that are expensive get quietly dropped in month nine.

Ollama · Groq · Codex · deterministic scoring · human-in-the-loop

A new project deploys with its guardrails already installed — five minutes in, the checks are running

One command scaffolds a repo, a Vercel deploy and an agent-ready kit, with a router that sends each job to the cheapest model tier that succeeds. Governance that is expensive to run stops getting run.

Next.js · scaffolding · agent kit · model-tier routing

Version-control record: 2026-05-21 → 2026-08-20 · 152 commits

It refuses to give medical advice, and the refusal is enforced in code rather than promised in a policy

A health dashboard with a vision model on top, and the project where I had to decide where a display tool stops and a regulated medical device starts. Under the Cures Act, software that only stores and displays device data is statutorily not a device; the moment it recommends a dose it is Clinical Decision Support. So four features were specified and then permanently killed, and a screen now runs over every word the model writes before anyone reads it. Separately, a model call per dish returns a carb and calorie range scored against a 5,800-image labelled dataset in real time and reported as mean absolute percentage error (MAPE) — it missed its own accuracy target, so the feature is switched off. The live app is not linked, by a standing privacy decision.

FDA device boundary · HIPAA scoping · dose-language screening · Groq · dataset-backed eval

Version-control record: 2026-05-21 → 2026-08-10 · 104 commits

Caught this portfolio's own demos overstating themselves — the audit tool turned on its owner

The tool I used to audit my own work — which is how several of the defects on this site were found, including demos that claimed to be live while making no model call.

Next.js · scoring rubric · audit harness

Screens a housing deal in under a minute and says go or no — not a spreadsheet to interpret

NorCal affordable-housing underwriting suite: HCV rent-cap screen, RCFE calculator, and deal screens that return a verdict rather than a number to interpret.

Next.js · TypeScript · real-estate underwriting

Packaging copy and product imagery for pennies, on your own machine — nothing uploaded, no per-use bill

Local-AI packaging studio: brand positioning copy and Stable Diffusion image prompts, with an SSRF-safe image API. Runs on local Ollama, so the marginal cost is zero.

Ollama · Stable Diffusion · Next.js · cost engine

One post becomes four channels overnight, and non-technical staff run it daily

One blog topic fans out to a Facebook caption, an Instagram caption with hashtags, and an email subject and body from a single model call. Runs as an admin tool used daily by non-technical staff.

Groq · Next.js · publish API · nightly cron

Version-control record: 2026-03-05 → 2026-08-10 · 187 commits

Shipped, live, and serving users on its own domain

Shipped and serving users. The demo runs the governance end to end on fictional families: consent that fails closed with teach-back and revocation, honest identity disclosure on a direct ask, and the deterministic family-alert ladder. The live product is not demonstrated directly, so this is the mechanism rather than described from guesswork.

Next.js

Version-control record: 2026-07-18 → 2026-08-20 · 433 commits

One set of questions answers the two things people actually need together — the tax bill and the benefits they qualify for

A wizard that collects filing status, income, household and deductions, then returns an estimate alongside benefits eligibility (including CalFresh) — the two questions people actually need answered together.

Next.js · multi-step wizard · eligibility rules

A couple sees the same picture of the week — load, plans, check-ins — instead of arguing from two different ones

The shared-home foundation for Calm Couch’s Our World: weekly focus and load tracking, family moments, structured check-ins, requests, agreements and repair flows. The demo runs on fictional data and ships no images; the live app is not linked, by a standing privacy decision.

Next.js · TypeScript · Supabase

Paying customers study for a licence exam in two languages — and no translated question ships until five checks agree it kept its meaning

Tutored exam-prep product with spaced-repetition scheduling and bilingual EN/ES content. The localisation pipeline is resumable and cost-capped, so a bad run is cheap to abandon.

Groq · Next.js · Supabase · TypeScript · Stripe · SM-2

Version-control record: 2026-02-18 → 2026-08-10 · 1,132 commits

You can argue with the inputs instead of having to trust the output — every projection shows the assumptions it stands on

A model whose inputs you can argue with. Numbers you cannot interrogate are numbers nobody should act on.

TypeScript · scenario modelling

Stops an inconsistent asset run before it reaches production — a second model checks the first and halts the line on disagreement

One model generates assets, a second verifies them, and the run halts if drift crosses a fixed threshold. It exists because a single model generating a large asset run drifts confidently and quietly, and nobody notices until the whole set is wrong.

GPT-4o · Gemini vision · consensus gate · spec-driven pipeline

Version-control record: 2026-04-09 → 2026-08-20 · 169 commits