DeepSense
Who: parents and guardians of autistic children aged 3 to 10 in Azerbaijan, and the specialists who work with them (speech therapists, ABA therapists, psychologists, special educators). What goes wrong today: 1. Between therapy sessions, parents cannot see objectively what their child is learning or struggling with. They rely on impressions, so problems such as a skill getting worse are noticed late. 2. The child's information is scattered. Diagnosis, co-occurring conditions such as ADHD or epilepsy, and doctors' opinions sit on paper in different places, and parents must re-explain everything to every new specialist. 3. Finding a specialist depends on word of mouth, with no easy way to compare specialization, experience, location or price. 4. Asking a general chatbot for help means pasting the child's name and medical details into it, and the answer is unchecked. It can give medication advice or invent numbers. What Learnly changes: the child plays short, calm lessons (hygiene, road safety, emotions) in a parent-controlled child mode, and every answer is measured. The parent sees clear progress (first-try accuracy, where the child gets stuck, 7-day trend) and gets a plain-language AI review with three home activities adapted to the child's sensory profile and interests. The AI never sees the child's name or medications, and every answer is checked before the parent sees it. When the parent is ready, they can send the full profile to a filtered, verified specialist with explicit consent, and access closes automatically when the collaboration ends.
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 testing17.0 / 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.
A finalist: the live jury scored it 76.0 / 100 on stage.
What the judges wrote
- GPT-6 AstraPreliminary81/100
- Strengths
- Your Learnly workflow connects lesson measurements, personalised parent guidance and consent-based specialist sharing, while keeping numerical calculations outside the model. Your documented testing includes concrete security checks, synthetic comparisons and failures with specific fixes, including model-overload recovery. Your running-cost estimates, synthetic-only demo and planned rehabilitation-centre pilot provide a credible path forward.
- To improve
- Your attached PDF describes FamCanvas.AI rather than Learnly; provide matching materials with an end-to-end example showing lesson completion, updated metrics, AI guidance and access revocation. Your evaluation should include a well-prompted, data-minimised baseline and specialist ratings of advice quality, not just compliance with your own guardrails. Your pilot should validate lesson content and escalation thresholds, and distinguish in-app accuracy from real-world skill gains. Your privacy tests should also challenge identifiers and medication details embedded in free text, beyond structured-field exclusion and name redaction.
- Claude Opus 5.5Preliminary80/100
- Strengths
- You define a sharp user (parents of autistic children aged 3–10 in Azerbaijan, plus their specialists) and concrete pains: no objective view between sessions, scattered records, and unsafe pasting into general chatbots. Your design rule that code computes the numbers and AI only interprets them is sound. It is backed by guardrails for medication advice, invented percentages, wrong language and fake lessons, plus a deterministic fallback report. Your testing is unusually thorough for a hackathon: 47 unit tests, 35 live row-level-security checks, a 12-profile eval against a baseline, and a candid table of what broke and how you fixed it, such as the Gemini 503 errors and the model fallback chain. Your feasibility section gives real numbers (about $0.01 per review, Supabase Pro cost), acknowledges the paid-tier privacy issue, and names a concrete pilot.
- To improve
- The pitch deck you uploaded is for a different project ('FamCanvas.AI' by Team Obsidian), so please make sure the right file is submitted for the final. Your baseline sends full raw data with a plain prompt, so the 0/12 vs 12/12 result mostly shows the value of data minimisation. A fairer comparison would be a well-prompted chatbot, or a therapist-written summary, judged on advice quality. Get even a handful of specialist ratings of the home-activity suggestions, because usefulness of the AI output is the claim least supported today. You submitted to the Enterprise track but the README says 'AI for Learning & Work', so tie the product to an organisational workflow and adoption metric, such as a rehab centre's intake or between-session monitoring.
The team
Fidan Jafarli
product
Yahya Jamalzade
developer
Rasul SadigliLead
data