LeadProof
USER: the head of sales at a residential real-estate developer in Azerbaijan. Managers take buyer calls in Azerbaijani, Russian and English, often mixed in one call. PROBLEM: Most calls are never checked: listening to all of them is not realistic, so teams review a small sample. What is found is found too late: weeks after the call, often only when a client complains. A missed call is costly. A manager promises what the developer cannot deliver (a bigger discount than allowed, an earlier handover date than in the project declaration, “guaranteed” price growth), and it comes back as a dispute, a refund demand or a legal and reputation risk. Or a hot lead gets no next step and goes to a competitor. Weak habits repeat because nobody sees them. WHAT LEADPROOF DOES: AI extracts the facts of each call with exact quotes. Our code checks every quote against the transcript and decides against rules loaded in advance: the company’s sales standards and the project declaration. The demo uses a basic rule set and a sample declaration; in real use the rules are set per developer and per residential complex. 6 sales-standard criteria (C1-C6); every manager promise checked against the sales rules and the declaration; a ready next action: correct the promise, call a hot lead back today, invite to the showroom, or coach the manager. The head of sales sees each finding with its quote and approves the action in one click. WHY IT IS DIFFERENT: 100% of calls, not a sample. Minutes after the call, not weeks: the audit takes ~6 s once the transcript is ready, so a broken rule is flagged while the manager can still fix it. Every finding is backed by an exact quote, and code, not the model, makes the decisions. No real quote, no fact: this blocks hallucinations.
Preliminary score
- Value for the user22.0 / 25
A specific problem and a clear outcome from solving it.
- Prototype and use of AI25.5 / 30
A working core scenario, and what AI actually contributes to the result.
- Quality testing18.0 / 20
Tests, examples of failures, and comparison with the current approach.
- Feasibility12.5 / 15
Data requirements, running costs and a clear next step.
- Originality7.0 / 10
A meaningful difference from common solutions.
A finalist: the live jury scored it 73.0 / 100 on stage.
What the judges wrote
- Claude Opus 5.5Preliminary86/100
- Strengths
- Your problem framing is precise: a head of sales at an Azerbaijani developer who hears under 1% of calls, with named risks such as guaranteed price growth or handover dates outside the project declaration, and a concrete next action for each finding. The architecture is the standout. The model only extracts facts with quotes, code verifies every quote against the transcript and makes all decisions, and a human approves, which gives a real defence against hallucination. Your testing is exemplary for a hackathon. You ran 40 labelled calls against a keyword baseline (40/40 vs 21/40), checked repeatability (230/230), and wrote a stress set after freezing the prompt, then honestly reported the S-09 speaker-label failure. You also documented operational breakages such as the 205 s call, the hung request and the UI silently showing mock data. Your unit economics are measured from real token counts rather than guessed, including the announced price increase.
- To improve
- All 50 calls were written by your own team with the same AI help that shaped the prompt, so the 40/40 result mostly shows internal consistency. Even a handful of anonymised real or role-played calls transcribed by actual speech-to-text would make the claim much stronger. Your README still says 22 test calls while the deck and submission say 40, so update it so reviewers can match the claims to the repo. Add a measured false-positive rate on clean calls. Also validate the market-research volumes and deal-saving assumptions with one developer, because the value case currently rests on unconfirmed averages. Moving the speaker-role check and transcript upload forward would turn the demo from fixed data files into something a pilot customer can try on their own calls.
- GPT-6 AstraPreliminary84/100
- Strengths
- Your focus on Azerbaijani real-estate sales leaders connects risky promises and missed follow-ups to specific, reviewable actions. Your combination of AI extraction, deterministic declaration checks and quote-linked human approval gives AI a meaningful but bounded role. Your reported repeated runs, frozen-prompt stress tests, keyword baseline and documented failures provide strong testing evidence, with transparent disclosure of synthetic data.
- To improve
- Your next validation step should be a blind pilot on consented real calls, independently labeled by sales leaders, measuring per-action precision, recall and review time saved. Strengthen speaker attribution and test whether verified quotes actually support extracted claims; exact text matching alone cannot prevent misinterpretation. Clearly distinguish cached fallback results from fresh audits and reconcile the README’s 22-call suite with your reported 40-call benchmark. Extend your deployment economics to include transcription and human review, alongside a concrete plan for access controls and call-data retention.
The team
Ilham Ağazadə
designer
Asim Mammadrzayev
developer
Murad RəhmanovLead
developer