Roadwise
Who: learner drivers in Baku preparing for the state driving exam, and the driving schools that train them. Today a learner gets a fine or an instructor's remark — it says what went wrong, never why, which mistake would fail the exam, or whether they're actually improving. What Roadwise does: the learner drives a 3D simulated city under international road rules, with fines based on Azerbaijan's Administrative Offences Code (İXM): ~1,000 road-sign poles placed by the standard sign rules, signalised junctions with stop lines, zebra crossings, give-way junctions and a 60 km/h town limit. After each drive they get a debrief tied to the exact time and street of every mistake, what to fix first, practice drills, a quiz on their weak rules, and an exam-readiness score tracked across drives. Why the assessment is trustworthy: the AI never judges driving. Rule checks, fines and all metrics come from deterministic, unit-tested code (304 tests); the AI only explains them. Every AI answer is checked against the drive data (event ids, rule keys, no invented numbers, every major fault covered) and replaced by a template if it fails. On our eval: 100% grounded, 0 hallucinations, 100% of major faults covered, $0.002 per debrief. Progress you can audit: readiness is a transparent formula (faults, rule compliance, speeding time, smoothness) with a point-by-point breakdown, not a black-box score. Honest limits and next step: the simulation has no AI traffic yet, and readiness hasn't been validated against real exam results. Our next step is a pilot with a Baku driving school to compare readiness scores with students' actual exam outcomes.
Preliminary score
- Value for the user20.5 / 25
A specific problem and a clear outcome from solving it.
- Prototype and use of AI23.0 / 30
A working core scenario, and what AI actually contributes to the result.
- Quality testing16.0 / 20
Tests, examples of failures, and comparison with the current approach.
- Feasibility13.0 / 15
Data requirements, running costs and a clear next step.
- Originality7.0 / 10
A meaningful difference from common solutions.
In the final, the live jury awarded this project 1st place with a score of 91.0 / 100.
What the judges wrote
- Claude Opus 5.5Preliminary80/100
- Strengths
- Your problem framing is specific and well sourced: Baku learner drivers, a practical exam where one major fault fails, and lesson and retake costs pulled from real local sources. The architecture is solid and honest. Deterministic, unit-tested code handles rule checks, fines and readiness, while Gemini only explains them, and a validator checks event ids and rule keys before falling back to a template. Your testing goes well beyond 'it works': 304 tests, an eval harness against a rules-only baseline, and concrete failures you found and fixed, such as quiz generation inventing numbers 21% of the time, plan retries at 33%, leaked internal keys and the 1 s cross-region latency. The full flow is real end to end, from the Unity game through device-link login and telemetry upload to a debrief in EN/RU/AZ.
- To improve
- Your AI evaluation runs on only 8 synthetic drive fixtures, so the 100% grounding and recall figures say little about messy real drives. Log real playtest drives from a few learners and report the metrics on those. Readiness is the headline user outcome but has not been checked against real exam results, so make the driving-school pilot concrete: how many students, what you will measure, and the threshold that would count as success. The deck still shows unfilled placeholders on the AI slide ('Grounding rate [__%]'), and the 'real city' claim is undercut by the lack of AI traffic. Close both gaps, or state plainly what the current map covers.
- GPT-6 AstraPreliminary79/100
- Strengths
- Your focus on Baku learner drivers connects a concrete practice-cost problem to locally relevant rules and event-specific feedback. Separating deterministic assessment from AI explanation, with validation and template fallbacks, is a sensible reliability design. Your reported quiz hallucinations and plan retries, supported by reproducible testing commands, provide specific evidence of failure discovery rather than generic success claims.
- To improve
- Strengthen your prototype evidence with an actual Unity-drive telemetry trace and its matching in-game/web debrief, clearly distinguished from synthetic fixtures. Report evaluation sample counts, first-attempt validity and fallback rates per language on unseen drives, and compare coaching usefulness against the rules-only baseline rather than counting tips alone. Have a local instructor verify provisional fines and exam-fault classifications. Use your planned school pilot to validate readiness against instructor assessments or exam outcomes, explicitly accounting for the current absence of AI traffic.
The team
Tural Gulumov
developer
Nijat Nasirli
developer
Nihad Gasimli
developer
Toghrul AbdullayevLead
developer