NeuroBridge
AI Enterprise SolutionsGPT & Claude

ClaimLens

Our primary users are motor insurance claims adjusters in Azerbaijan. They must review vehicle photos, identify visible damage, prepare repair estimates, and decide which claims need further inspection. Poor-quality photos and missing views create additional work and customer follow-ups. They need a faster way to prepare consistent, evidence-backed assessments while retaining control over the final decision.

VideoGitHub

Preliminary score

75.0/ 100Strongaverage of 2 judges
  • Value for the user19.5 / 25

    A specific problem and a clear outcome from solving it.

  • Prototype and use of AI24.0 / 30

    A working core scenario, and what AI actually contributes to the result.

  • Quality testing14.5 / 20

    Tests, examples of failures, and comparison with the current approach.

  • Feasibility11.5 / 15

    Data requirements, running costs and a clear next step.

  • Originality5.5 / 10

    A meaningful difference from common solutions.

What the judges wrote

  • Claude Opus 5.5Preliminary76/100
    Strengths
    Your prototype runs the whole adjuster workflow end to end. It covers photo quality checks (Laplacian blur, coverage gaps), Gemini vision for vehicle ID and damage boxes, a rule-based itemised estimate priced from live marketplace listings with links, fraud signals, and an adjuster queue. Your split between AI perception and deterministic, explainable cost and recommendation code is a sound design for insurers who need reviewable decisions. Your failure log is unusually honest and concrete: the double-charged headlight, the 70 AZN hood-cover mismatch, the Lada and E-Class catalog gaps, and the stuck-processing bug. Those fixes sit alongside 113 automated tests and Playwright runs.
    To improve
    Your deck screenshot shows a $13,562 estimate for a 1987–2004 Lada Samara, built almost entirely from placeholder fallback prices. Add a total-loss or vehicle-value sanity check, and replace placeholder labour and paint rates with local AZN rates as early as possible. Measure damage-detection accuracy on even 20–30 hand-labelled photos, and time one adjuster doing a claim manually versus with ClaimLens, so your 50% time-saving target has a baseline. Your running-cost figure is based on Claude Opus, but your live analyzer is Gemini 2.5 Flash; cost the model you actually run, including image tokens.
  • GPT-6 AstraPreliminary74/100
    Strengths
    Your workflow combines photo-quality checks, evidence-linked vision findings and explicit pricing rules while keeping the adjuster in control. Your testing report is concrete: duplicate headlight billing, incorrect accessory matches and restart recovery each led to targeted fixes. Your labelled fallback estimates, explicit demo mode and candid list of untested scenarios make the prototype's limitations clear.
    To improve
    Your next validation should measure missed damage, box accuracy, estimate error and preparation time against expert assessments and manual adjusting. Replace placeholder labour and paint rates with verified local quotes, and validate part fitment by year, trim and side. Add authentication and approval permissions before an insurer pilot. Reconcile the README's local-marketplace pipeline with the submission's stated replacement of that source, and measure actual Gemini-plus-Firecrawl costs and latency.

The team

  • Ughur MammadovLead

    product

  • Bayram Aliyev

    data

  • Nadir Askarov

    developer

  • Fazil Behbudov

    developer