Record video of your building. Get a professional ADA compliance report.

Passline is an ADA (Americans with Disabilities Act) compliance audit tool built to streamline accessibility inspections for buildings and commercial properties. Traditional ADA audits require expensive consultants and time-consuming on-site assessments

Passline lowers that barrier by guiding auditors through a simple, video-based inspection workflow.

Auditors start on a dashboard where they can create a new audit or review past ones. To begin, they enter basic property information, building type, year built, and available features like parking and elevators. The tool then prompts the auditor to upload video walkthroughs of the property. Passline automatically classifies each space from the footage — identifying areas such as entrances, restrooms, corridors, parking lots, service counters, elevators, stairs, ramps, and signage, with no manual categorization required.

Each classified space is analyzed against ADA standards. The system flags violations and warnings with specific ADA code citations (e.g., ADA Standards §404.2.3), estimated remediation costs, and detailed descriptions. Finally, a generated compliance report summarizes the findings with an overall compliance score and total cost range for remediation — printable as a PDF.

cd backend

# Create virtualenv
python3 -m venv venv && source venv/bin/activate

# Install dependencies
pip install -r requirements.txt

# Install FFmpeg (required for video processing)
# Ubuntu: sudo apt install ffmpeg
# macOS:  brew install ffmpeg

# Install DepthAnything V2 (optional — falls back to placeholder if unavailable)
git clone https://github.com/DepthAnything/Depth-Anything-V2
pip install -e ./Depth-Anything-V2
# Download ViT-S checkpoint:
# https://huggingface.co/depth-anything/Depth-Anything-V2-Small/resolve/main/depth_anything_v2_vits.pth
# Place at: backend/checkpoints/depth_anything_v2_vits.pth
mkdir -p checkpoints
curl -L -o checkpoints/depth_anything_v2_vits.pth \
  https://huggingface.co/depth-anything/Depth-Anything-V2-Small/resolve/main/depth_anything_v2_vits.pth

# Set environment variables
cp .env.example .env
# Edit .env with your API keys

# Run
source venv/bin/activate && uvicorn main:app --port 8000

cd client
brew install node
npm install
npm start   # Runs on http://localhost:3000

MONGODB_URI=mongodb+srv://...
MONGODB_DB=passline
GEMINI_API_KEY=...
FEATHERLESS_API_KEY=...
FEATHERLESS_MODEL=meta-llama/Meta-Llama-3.1-70B-Instruct
UPLOAD_DIR=./uploads
FRAMES_DIR=./frames
REPORTS_DIR=./reports
DEPTH_MODEL_PATH=./checkpoints/depth_anything_v2_vits.pth

User records video
    ↓
Frontend (React) → POST /upload → FastAPI
    ↓
FFmpeg: extract frames at 2fps
    ↓
OpenCV: SSIM deduplication → 3-20 key frames
    ↓
OpenCV: Canny + contour → reference object calibration (pixels/inch)
    ↓
Gemini 2.5 Flash: room classification (34 room types)
    ↓                                      ↓
Gemini 2.5 Flash:              module-level ADA rule set
universal feature detection    (Pass B — driven by room type)
(40 feature types)
    ↓
feature-level ADA rule set
(Pass A — driven by detected features)
    ↓
Violations from both passes merged + deduped
    ↓
DepthAnything V2 ViT-S: metric depth maps + measurements
    ↓
Featherless LLM (Llama 3.1 70B): narrative descriptions for each violation
    ↓
ReportLab: PDF generation
    ↓
MongoDB: audit document storage
    ↓
Frontend polls /status → displays report

The pipeline sends the same video frames to Gemini twice for two independent signals:

1. Feature detection — identifies every ADA-relevant element visible in the video (doors, grab bars, ramps, signage, etc.) across a 40-type closed enum. Each detected feature is evaluated against feature-level ADA rules (Pass A).

2. Room classification — identifies the room type from the same frames (one of 34 types: restroom, hallway, parking, elevator, etc.). The classified type drives a separate set of whole-room ADA rules (Pass B) covering requirements that may not be visible as individual features — for example, a restroom classification triggers required grab-bar checks even if no grab bar appears in the video.

Violations from both passes are merged and deduplicated into a single list per room.

1. Best: Reference object detected → pixel calibration → accurate inch measurements
2. Good: No reference, metric depth model → flagged as "estimated"
3. Fallback: Gemini relative classifications only → flagged as "screening estimate"
4. Demo floor: Feature-presence rules only (no measurements needed)

Anthropic Claude Code (Opus 4.6) was used for integration of front and backend, construction of README.md and SPEC.md, backend construction, and various bug fixes.