Skip to Main Content

Team Abrar

Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

Team Abrar — Demo video

Getting a commercial or residential building permit takes anywhere from 4 to 16 weeks per review cycle at city planning departments. Projects lose thousands of dollars each day while drawings sit in administrative backlogs. At the same time, overworked human examiners regularly miss accessibility and egress violations, such as restroom doors that swing inward and block the required wheelchair turning circle under ADA Section 604, or school corridors that fail minimum egress clear widths under IBC Section 1020. Catching these problems weeks into the process forces architects to redraw the plans, resubmit, and restart the waiting game. We built PlanPass to compress this pre-check and remediation process from weeks into about 30 seconds. An architect or city examiner loads a CAD floor plan or picks from our industry benchmarks (schools, hospitals, industrial plants, cafes, and residential ADUs). The reviewer can speak naturally through a push-to-talk voice bar to set the audit scope. PlanPass then runs a closed-loop inspection: 1. Visual Workplane Inspection: Instead of relying on brittle bounding-box segmenters, the model inspects the drawing's linework directly, identifying doors, corridors, fixtures, and clear dimensions from CAD pixels. 2. Live Statutory Grounding: The agent queries live federal rules from the US eCFR REST API (Titles 36 and 28 for ADAAG) and geocodes the property address using OpenStreetMap Nominatim to resolve local city, county, and state jurisdictions. 3. Deterministic FastMCP Math: Physical clearances are calculated against IBC 2024 and ADA 2010 rules. Unverified citations are rejected rather than guessed. 4. Closed-Loop Remediation: If violations exist, PlanPass generates both a redline markup for drafters and a clean revised plan with fixes baked in. It then re-audits the clean drawing to unlock 100% Professional Engineer permit approval, stamped PDF packets, and sovereign JSONL training datasets.

Team