Helpious AI

Helpious AI — A photo or a voice note becomes a structured job brief for the homeowner, and the tradie's quote is drafted only from their own price list. Live product: https://helpious.com Code (hackathon work only, public): https://github.com/alihassan1437/helpious-hackathon-ai Disclosure: Helpious is a live Australian tradie marketplace that existed before this hackathon. Only the AI features below — built 26 Oct to 4 Nov 2026 — are claimed as hackathon work. What we built, all through one gateway (POST /api/ai) with login and role checks, a global kill switch, personal-data stripping before any model call, per-user daily caps and clean fallbacks: - Snap-to-Quote: a photo plus the customer's words become a structured brief (trade, issue, urgency) that pre-fills the quote form, editable. - Voice briefs and voice quote drafts (Muse Voice Transcribe): transcript shown and editable; spoken phone numbers stripped. - Lead brief: a three-line AI brief per job, generated once, then cached. - Quote draft: line items only from the business's own price list, enforced twice (prompt plus a deterministic check); a price that is not theirs is left blank, never invented. The AI never sends anything. - Paste-to-profile: a draft listing built from pasted text, every field shown against its source sentence, nothing saved without a tick. - Fixed safety card on dangerous jobs (gas smell, sparking, flooding near power): written by us, never AI-generated; never blocks the job. Models: Muse Spark 1.3, Standard tier only (the Contributor tier is rejected in code); Muse Voice Transcribe 1.0; SAM 3.1 integrated and API-verified, with photo analysis deliberately off until on-device masking ships — briefs come from the customer's words, and the app says so. Evaluation: 112 cases (100 jobs across 10 trades plus 12 adversarial) run twice against the live gateway. Run 1 failed honestly (trade 26/100, safety 4/6); we fixed the causes and re-ran identically. Run 2: trade 94/100, urgency 87/100, valid output 112/112, safety recall 6/6, personal-data leaks 0, invented facts 0. Stated plainly: median response 6.4s misses our own 4s target (p95 9.9s; repeat views cached at ~0.6s). Total Meta Model API spend for the whole hackathon was about US$2 of the $150 credit. Demo access (test data only): sign in at https://helpious.com/login — Judge (customer): demo.judge@example.com; Business: demo.business@example.com; Plumber: demo.plumbing@example.com. Passwords — Judge: MUkPMA4vIXrblW23MDrhyLrD; Business: BizDemo!3pW8sT5kYrV2; Plumber: fInLVsGh6J/lzmbPujY89+M6. Walkthrough: as judge, open Saved businesses, then Darwin Demo Plumbing; go to Get a quote and describe a problem by text or voice; describe a gas smell to see the safety card. As the business, open the waiting Casuarina lead for the AI brief and Draft with AI.