Circular Capture Coach

💡 Inspiration In circular supply chains and industrial e-scrap management, accurate visual asset recording is the bottleneck to efficient sorting, valuation, and recycling pipelines. Field technicians frequently submit device captures suffering from motion blur, blinding glare, cut-off equipment edges, or completely obscured tracking tags. This introduces heavy administrative friction, causing thousands of hours of manual re-inspection cycles.

Circular Capture Coach solves this problem directly at the source. It functions as an automated visual compliance assistant, analyzing photographs in real time to provide immediate quality scoring and clear, step-by-step camera re-shoot coaching.

⚙️ What It Does Circular Capture Coach is a comprehensive visual compliance workspace optimized for sorting operators. The application automates two distinct evaluation layers natively:

Set-Level Checklist Ingestion: It parses official challenge data tracking sheets (photo_sets.csv), grouping unique Set IDs and device relationships to calculate missing mandatory views (front, rear_ports, label).
Granular Visual Quality Audits: It uses multi-modal AI to audit raw pixel matrices across five strict defect parameters (blur, underexposed, glare_or_overexposed, framing, label_obstructed), cleanly separating poor-quality photos from completely missing views.

The engine instantly returns clear status badges (USABLE, RETAKE, HUMAN REVIEW), formats multiple errors using a proper semicolon-separated logging format, isolates defects from black data-privacy masks, and generates precise camera alignment guidelines to coach users on how to correct their shots. Technicians can then instantly download a structured text manifest log file for local enterprise record-keeping.

As an optional track extension, the platform features a fully responsive, un-mirrored "Real-Time Live Camera Mode" that activates a device's lens, allowing workers to execute instant visual compliance assessments right on the warehouse floor.

🧮 Theoretical Framework & Quality Metrics To validate visual compliance across an asset cluster, the platform cross-references each image frame against a multidimensional mathematical grading matrix. We model the overall set compliance score C as a joint probability configuration across the four primary computer vision inspection layers:

Where each individual metric component maps directly to a discrete operational evaluation layer:

Focus Threshold Layer (P(Blur)): Evaluates high-frequency spatial variation to identify pixel-level motion blur or focus drift.
Illumination Uniformity (P(Lighting)): Evaluates local contrast ratios and pixel intensity variations to isolate underexposed or glare_or_overexposed conditions.
OCR Tag Legibility (P(Label)): Validates bounding box text clarity across critical manufacturer serial markers, ignoring intentional privacy blackouts.
Spatial Geometry (P(Framing)): Measures aspect ratios and device centering thresholds relative to the camera viewport boundaries.

If any isolated system parameter drops below acceptable thresholds, the overall compliance maps to zero:

The engine automatically flags the entry as RETAKE, logs the specific defect codes using the required semicolon-delimited schema, and auto-generates localized camera re-shoot instructions.

🛠️ Technical Implementation & Architecture The prototype is engineered using a lean, decoupled Python architecture optimized for execution speed and developer iteration agility:

Frontend User Space (Streamlit): Provides a fluid, mobile-responsive layout workspace. Technicians can toggle between batch CSV analysis and direct webcam capture views.
Data Integration Pipeline (Pandas): Ingests tabular challenge manifests natively, indexing image lists dynamically by matching multi-case extensions (.jpg, .png, .JPEG) across local repository directories.
API Integration Middleware (google-genai SDK): Safely bundles raw image byte arrays with structural instructions and routes them across a secure, authenticated link.
Multimodal Engine Core (Gemini 3.6 Flash): Processes visual pixels and structural instructions simultaneously, parsing response streams into strict JSON text blocks to drive the user interface.

🧠 Challenges We Ran Into & Workarounds Building a high-throughput parallel execution engine threw network rate limits due to the free-tier quota cap of 20 requests per minute. We successfully bypassed this barrier by engineering an automated paced loop system with an integrated 1.5-second sleep step, clearing API ceiling thresholds perfectly without losing processing integrity. Additionally, resolving browser layout mirroring properties required deploying custom CSS override wrappers to align right-to-left user hand motions natively.

🏆 Accomplishments We're Proud Of We successfully developed a completely functional, dataset-agnostic production prototype entirely from scratch as a solo developer inside a tight timeline. The platform successfully automates compliance tracking across all 9 official challenge data sets, handles complex multi-image layout columns fluidly, and bridges the gap between raw pixel analysis and clean structural business diagnostics.