Məskən — AI decision engine for choosing where to live in Azerbaijan
User: a first-time home buyer in Baku — typically a young family choosing where to buy a flat, a house or a small commercial unit, with a fixed budget and needs like "near the metro" or "near a school". Problem today: listing sites (bina.az, tap.az) show individual flats, not neighbourhoods. To judge an area the buyer has to piece together Google Maps, word of mouth and guesswork: how far the metro really is, which schools, shops and pharmacies are nearby, whether roads are paved, what prices are normal there. This takes hours to days and still ends in a gut decision. Prices make mistakes costly: a 65 m² average Baku flat costs ≈178k AZN, and new-build prices are rising faster than wages. Legal risk is hidden in plain sight: in our aggregate of Baku listings, 21% state the property has no title extract ("çıxarış yoxdur"), and buyers often don't know what phrases like "etibarnamə ilə" mean. Outcome with Məskən: the buyer types one sentence (e.g. "Nərimanovda metroya yaxın ev lazımdır") and in about two minutes gets 5 ranked neighbourhoods on a map with metro walking distance, nearby services, price level, a computed confidence score, and a red-flag check of the listing text.
Preliminary score
- Value for the user21.0 / 25
A specific problem and a clear outcome from solving it.
- Prototype and use of AI25.0 / 30
A working core scenario, and what AI actually contributes to the result.
- Quality testing17.0 / 20
Tests, examples of failures, and comparison with the current approach.
- Feasibility13.5 / 15
Data requirements, running costs and a clear next step.
- Originality8.0 / 10
A meaningful difference from common solutions.
A finalist: the live jury scored it 77.0 / 100 on stage.
What the judges wrote
- Claude Opus 5.5Preliminary85/100
- Strengths
- Your user is specific (a first-time buyer family in Baku on a fixed budget), and the pain is backed by sourced numbers: price growth versus wages and the 21% of listings with no title extract. Your pipeline is cleanly designed: Gemini parses Azerbaijani intent and writes grounded explanations, while deterministic code ranks the areas and computes the confidence score. That makes the AI essential but keeps it accountable. Your evals are a standout. You compare the AI against a keyword baseline, report precision and recall honestly (86% vs 100% precision), measure grounding at 638/639 numbers, and give a detailed failure analysis with root causes and fixes, such as Xankəndi's pre-2023 OSM data and the phantom metro stations. Feasibility is also concrete: measured token costs per session, data sources labelled by status, demo values flagged, and a 30-day realtor pilot.
- To improve
- Your user outcome (a shortlist in about 2 minutes, compared with hours or days today) is still asserted. The README leaves the user timing test as a TODO, so running even 3 buyers through bina.az plus Google Maps versus Məskən would give you the comparison with today's approach that your testing lacks. Your eval sets are small (10–30 cases), and the legal check has not been reviewed by a lawyer. Expanding both, and having a qualified reviewer validate the risk phrases, would make the 100% recall claim much more convincing. Given the Enterprise track, name the single adoption metric a realtor or bank would track in the pilot, for example time-to-shortlist or lead conversion. Also reconcile small inconsistencies across your materials: confidence is 69 in the deck but 60 in the README, and the deck says 1,125 neighbourhoods where the README says 3,700 micro-areas.
- GPT-6 AstraPreliminary84/100
- Strengths
- Your first-time buyer scenario addresses a concrete, costly decision with a clear shortlist outcome. Your architecture sensibly separates AI intent parsing and explanations from deterministic ranking, constraint checks and confidence calculation. Your baseline comparisons, documented failure fixes, source labels and token-based cost estimates provide strong technical substance beyond a typical property chatbot.
- To improve
- Your planned buyer/realtor pilot should measure time-to-shortlist and shortlist usefulness against listings plus maps; the two-minute benefit remains unvalidated with users. Expand testing with held-out queries, lawyer-reviewed legal examples and impossible constraints, clarifying whether the README’s relaxation step can override hard requirements. Check that explanations attach numbers to the correct facts and that confidence scores reflect recommendation quality, rather than relying on numeric presence alone. Reconcile the deck and README’s area definitions and counts, and make price-sample age and walking-distance approximations explicit beside recommendations.
The team
Sarkhan Mahabbatli
developer
Samir HasanovLead
data
Nijat Shabanov
developer