JACQUES Agentic OS

Miles Tracker

The phone app a driver runs. Scan a QR code to join an agency, tap once to share a time-limited link, and the phone becomes a GPS tracker. It is a Flutter fork of the open-source Traccar client with the Firebase telemetry removed.

1

tap to share a link that expires when you say so

0

third-party trackers in the app

Built with Flutter and Dart; iOS build proven through CI from a Windows machine; Android APK side-loadable by QR before either store listing exists.

TripPulse

A traveler's co-pilot for the moment a trip breaks. It ranks air, rail, road and transit options by speed, cost, reliability, accessibility and carbon, tells you which refunds the 2024 US DOT rules owe you, and drafts the claim.

Not public yet. The code is one deploy from live.

249

transportation data sources in its library94 of them real-time

6

live federations: flights, trains, airspace status, weather alerts, directions and transit

Built with A no-build progressive web app; Cloudflare Functions and KV; FlightAware, Duffel and Amtrak feeds; web push; a GitHub Actions sidecar that pre-warms flight prices.

50-state agency study

Before building Miles Tracking I measured the market. A crawler read every US transit agency website it could find and recorded whether riders can see their bus today.

8,167

pages read across 787 agency websites

1,488

agencies with no rider-facing tracker747 had one; 409 were marked strong prospects

Built with Python, with the findings in a spreadsheet and an HTML report.

Amtrak Operational Intelligence

One self-contained HTML file that joins twenty-eight months of train status with the full timetable corpus and a live position feed. On-time performance comes with a confidence interval, and the Pain Points tab ranks recurring problems by frequency times severity.

3.2M

station events across 584 trains and 533 stations

576

timetable PDFs parsed into 364 trains and 49 routes

0

network requests needed to open it

Built with React, Recharts and Leaflet in one 12.7 MB file; a compressed data cube; Python for the extract and transform.

Diamond Dashboard

A daily screener for the options wheel. Each morning it pulls chains, earnings dates and the macro calendar, scores every candidate on nine components, and writes one offline HTML sheet. A pre-registered study records whether the score predicts anything, in an append-only ledger.

246

unit tests across the pipeline

9

score components, each explained on the sheet

Built with Python standard library only; a Windows scheduled task; a public sheet with no private data on it.

CDAN-NG crash data networkAt work

A clean-room prototype of a modernized national crash-data network. It takes crash feeds from five states in five formats and maps them to the national standard, and it is built around the error a matching field name hides: the names agree but the definitions do not. Every record ends in exactly one state, lineage is traceable, and an AI suggestion applies only after a named person approves it.

Built for a federal request for information. Every crash record is simulated; the one real mapping comes from a published state report.

204

automated checks: 121 on the engines and exports, 83 in a real browser

50,000

records in the scale run

866

national attributes mapped for one real state crosswalk382 unmapped; 303 of them need a change to the state's crash-report form

Built with React 18, TypeScript, Vite, Tailwind and Recharts; pure TypeScript engines; Cloudflare Pages Functions or a Node container with a PostgreSQL schema; Playwright; a one-file offline build that makes no network request.

FAA Operational IntelligenceAt work

One file over seven years of US flight operations at the top 200 airports, built to answer where the system breaks, why, and what to fix first. It runs from an overview through pain points, root cause, a drill-down map and customer impact to a ranked list of actions.

42.1M

flights tracked, 2019 to 202544.0M in the data window

934K

cancellations

15.9%

of carrier and airport clusters drive 75% of delay minutes

Built with React and Recharts in one 7.4 MB HTML file; Python for the extract and transform; public on-time, capacity and safety-report data.

MetroPulseAt work

A performance dashboard for the chief performance officer of a large metro agency, built to replace the seven tools opened every morning with one screen. Thirty-five months of ridership sit beside live incidents, vehicle positions, arrival predictions, elevator outages and the weather.

A proposal prototype on the agency's public data. The agency is not named here.

35

months of ridership, January 2023 to November 2025

6

views, from executive summary to rider impact

30 s

refresh on the live panels

Built with A single-file HTML dashboard; a small Node proxy for the agency's public APIs and the weather service.

TransitPulseAt work

Why buses run late, explained with the data a city already publishes. On-time performance, ridership and the schedule are joined with 311 complaints, street closures, crashes and emergency dispatches, so a recurring delay shows its cause beside it.

A concept built on open data, not endorsed by the agency, which is not named here.

104

routes, 6,428 stops and 1.27M scheduled stop times

929,773

311 service requests joined to the network

65.7%

weekday on-time performance, matching the agency's published figure

Built with A single-file HTML dashboard with Chart.js; Python for the analytics; a SQL schema over the city's open-data portal.

Warranty RecoveryAt work

Finds money a fleet is already owed. It connects to the maintenance system and looks for three things: warranty claims never filed, repairs that match a later recall campaign, and vendor penalties assessed but never collected. Each finding carries an evidence trail and a drafted claim package.

No dollar headline is quoted here on purpose: the first one was an artifact of a single rule, and the project retracted it.

210

automated tests, green on 19 September 2026

943,586

warranty coverage rows in the real datasetwith 1,145 claims and 1,554 penalties

3

kinds of finding, each with its evidence

Built with Python, DuckDB and Streamlit; pytest; Docker; readers for public recall data from NHTSA, Transport Canada, the FAA and the FRA.

Sentinel HSE

Incident reporting for oil and gas field workers in five languages, including Hausa, Yoruba, Igbo and Nigerian Pidgin. A deterministic escalation rule runs before any AI sees the report, and the assistant either cites the safety document or says it cannot answer.

5

languages with zero untranslated strings

370

unit tests, plus 47 smoke, 125 wiring and 22 blocker checks

16

guarantees enforced in the database itself

Built with React 19 and TypeScript; Cloudflare Functions; Postgres with pgvector; Docker for the full stack; Vitest and Playwright.

Lead Snap

Photograph a business card and it becomes a contact, a research brief and a row ready for Salesforce, HubSpot, Zoho or Dynamics. A guest mode runs the text recognition entirely in the browser so nothing leaves the phone until you choose.

Not public yet.

27

parser cases, every one green

4

CRM export presets, written without a spreadsheet library

Built with React and Vite; Cloudflare Functions, D1 and R2; Workers AI for vision OCR and briefs; Tesseract.js for the offline guest mode.

DATA-TALKAt work

A natural-language assistant over enterprise data: a business user asks a question in plain English and gets the answer, the SQL and a chart. My part was the demo film: re-cut, re-narrated with a local neural voice and given a sound design, so the product shows itself in under two minutes.

164

of 181 evaluation questions answered correctly in the published harnessvisualization 21 of 21; plain SQL 29 of 30

104 s

film, cut from 120, mixed to broadcast loudness

29

synthesized sound cues

Built with For the film: Python with ffmpeg, Demucs for stem separation, Kokoro text to speech and Whisper for the transcript.

Fleet SentinelAt work

A predictive-maintenance layer for transit and municipal fleets. A failure model scores every component and rolls up to an asset health band; the system drafts the work orders and the federal asset-management narrative, matches grants to programmes, and answers questions through an agent with tools over the maintenance database.

5

role views: executive, dispatcher, asset, compliance and grants

18

tools the agent can call over the production data

65

production tables ingested

Built with Single-file React dashboards; Python with DuckDB, Streamlit, Weibull survival models and LightGBM; an agent with tools; a GTFS-realtime poller writing to object storage.

DOT concept prototypesAt work

Ten clickable concepts for federal transportation agencies, each one file. One Voice DOT is the real one: ask a travel question and get one plain-English answer with citations, pulled live from seven agency feeds. Evacuation Conductor orchestrates a multi-modal evacuation; a workforce twin simulates a retirement cliff; a response engine drafts cited answers to public requests.

10

concepts, each opening as a single file

7

federation adapters behind One Voice DOT, with a health probe

576

timetable PDFs parsed into 364 trains and 49 routes for the rail answers

Built with Single-file React bundles for the concepts; One Voice DOT is Next.js 14, TypeScript and Tailwind with server-side adapters for each agency feed and an optional model step that rewrites the answer in plain English.

Data Agents and LibrariesAt work

A catalogue of catalogues: every outside data source a team might need, found by parallel research agents and written into one 22-column record with a stable id. Transportation, test data, real estate, youth sports, fundraising and transit agencies, searchable in one offline page.

4,572

records across 9 libraries2,340 data sources, 2,335 unique; the rest are directory records, counted separately

320

sources gated behind an approval list; nothing gated was downloaded

65

browser assertions pass, with contrast checked on 21 colour tokens in both themes

Built with Python and Node build scripts; CSV and YAML as the source of truth; generated single-file dashboards, light and dark; a Playwright gate.

Fun Finder

A nationwide search for things to do with your kids: sports, camps, parks and days out. Prices are shown only when they are known, accessibility filters are first-class, and children's profiles are nickname-only.

163k

programs across every US state

1.13M

parks, trails and places to go

4

languages

Built with React and Vite; Cloudflare Pages Functions and D1 with spatial search; sitemaps generated at the edge; a Claude Haiku fallback for queries the parser cannot read; Node pipelines over OpenStreetMap, Wikidata, the National Park Service and Eventbrite.

Receipt Snap

A household receipt app. Photograph the receipt, the amount and store are read for you, and the week's total and what is owed back are always one glance away. Unknown values are left blank rather than guessed.

234

assertions across five phone sizes and two browser engines, zero failures

Built with A progressive web app on Cloudflare with Workers AI reading the receipts.

ScanKit

An installable barcode and QR scanner for phones with four jobs: scan, keep an inventory, look an item up, and tick off a checklist. It runs from a browser and works offline once installed.

Not public yet.

4

jobs, one install

Built with React and Vite; the ZXing decoder in the browser; a service worker for offline use.

Fun Things database

A hand-verified set of at-home activities for kids, each with its source checked and a safety review. The build refuses to publish if any record breaks a rule.

134

fields in the schema

177

source links checked; 115 answered, 0 failed, the rest were not web pages

Built with Python; JSON shards compiled to JSON, CSV and a spreadsheet.

File FinderAt work

A search page for a company's network shares. Staff type part of a file name and see which folder it lives in. The index holds metadata only, never a file's contents, and the whole thing is one HTML page that works offline with nothing to install.

Shown in words only: a capture would show the company's own file names.

180,788

files indexed across 7 shares

0

dependencies beyond Python's standard library

1

file to open, about 13 MB, no server

Built with Python 3 standard library for the scan; one self-contained HTML page built from a template; a one-click refresh.

Time KeeperAt work

A cleaner replacement for a weekly timesheet page. A keyboard-driven grid, a required reason for every logged hour, autosave with retry, a warning as a task nears its funded ceiling, submit and recall, and copy last week. On a phone it becomes a day view with voice-to-text for the reason.

1

file to swap the demo data for the real timesheet API

2

layouts: a desk grid and a phone day view

Built with React 18 and strict TypeScript; Vite with a single-file build; Cloudflare Pages.

Narrated strategy deckAt work

A strategy presentation as one HTML file that narrates itself: slides, a guided tour, a narrator you can switch on, playback speed, and a menu, with the audio embedded so it opens anywhere with no player. Built to argue that a proposal should show the work instead of describing it.

An internal deck, so shown in words only.

1

file with the slides, the narration and the tour inside it

10

concepts argued end to end on one live pursuit

Built with A single HTML file with embedded narration; Python and Node for rendering, narration and screenshots; Cloudflare Pages.

Universal car fob

A feasibility study for one key fob that works across several cars. Four architectures compared, nine decisions recorded, nine bench tests designed, and one open question that blocks the next step. Nothing has been bought and nothing is connected to a vehicle.

4

architectures compared; one chosen

$0

spent

Built with Documents, with every claim pointing at a numbered source.

Prediction-market bots

Paper-only bots that compare sharp sportsbook consensus with prediction-market prices and size bets by a quarter Kelly. They prefer no trade to a bad trade, and a separate study showed a language-model forecaster could not beat the crowd.

314

tests across the two bots

11

candidates from one scan of 643 markets

Built with Python and pytest. Live trading is disabled in code, not in a setting.

Strategy Models

A multi-scanner paper trading system with hard caps, a market-breadth gauge and a review gate every backtest has to pass before anyone is allowed to believe it.

0.94

correlation between the breadth gauge and the S&P 500

6

model specifications, about 43 test files

Built with Python; daily market data from several providers; a local control room.