PITT - Driver-First Trip Assistant

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for PITT - Driver-First Trip Assistant

Inspiration

Delivery drivers must reconcile fuel reserve, detour time, delivery windows, changing conditions, and exception paperwork—often from incomplete information. PITT explores a driver-first decision layer: show the facts, explain a bounded option and its limits, and leave the decision with the driver.

What it does

PITT is a runnable local proof of concept with four seeded outcomes. It orders a delivery ledger, compares route options, checks predicted traffic, moving weather patterns, and a road-work register, and weighs simulated fuel prices against detour cost. Trip Watch lets the driver refuel, pass a stop, continue safely after a morning fill, or confirm an offered early closure when fuel reaches zero. A review-gated report and local machine handoff record the result without contacting external systems.

The demo clearly labels its boundaries: it uses deterministic seeded data, invented coordinates, and simulated prices. It does not claim live GPS, mapping, traffic, weather, construction, telematics, dispatch, or vehicle control.

How we built it

Codex and GPT-5.6 were central to the integration. Codex organized collaboration lanes and contracts, implemented and connected deterministic modules, added regression tests, and maintained evidence linking claims to code and validation. Browser testing exposed an important missing economic branch: cheap fuel is not worthwhile after a morning fill when the detour costs more than the pump saving. That observation became an explicit rule, UI explanation, unit coverage, and one of four end-to-end outcomes.

The published demo intentionally uses its deterministic local fallback, so judges can run it without credentials. A provider-neutral adapter preserves deterministic facts as authoritative if an approved AI narrative is added later.

Challenges

Keeping the project runnable without credentials, a database, or a live provider.
Demonstrating useful route and fuel reasoning without presenting seeded data as live fleet information.
Designing optional AI narration that cannot rewrite calculations or bypass driver review.

Accomplishments

27 Node tests for planning, moving weather, work-zone checks, Trip Watch, fuel consequences, report gating, and UI wiring.
Four Playwright runs proving every fuel-decision outcome shown in the demo.
24 Python tests covering fallback, provider failure, contract shape, and secret-safe configuration.
A complete driver-controlled flow from plan to exception review, outcome recording, and local handoff.

What we learned

A recommendation becomes more useful when its source, trade-off, and limitations are visible. The promising product question is not autonomous routing; it is whether an understandable, reviewable assistant can reduce decision and reporting friction for drivers.

What's next

Validate the companion with a small fleet, identify report fields that save time without duplicating paperwork, and design bounded adapters to systems customers already use.

Try the live demo · View the source