System Architecture & Serverless Pipeline

Architectural Principles

BestShots.AI is engineered to handle shoots of 5,000+ camera RAW frames without introducing client browser stutter or exhausting cloud function timeouts. To achieve this performance, the platform separates user-facing interaction from heavy computer vision workloads using an asynchronous, event-driven topology.

---

High-Level Topology

The system comprises four primary layers:

1. Client Tier (Next.js 15 App Router): Delivers 0ms optimistic UI updates via TanStack Query and Zustand. Handles client-side EXIF parsing and communicates directly with cloud storage.
2. Database Tier (Neon PostgreSQL with pgvector): Houses relational project metadata, audit logs, and 512-dimensional biometric embeddings. Connected via pooled and unpooled connection strings.
3. Storage Tier (Microsoft Azure Blob Storage): Stores original camera RAW payloads, compressed WebP proxies, and cropped face avatars behind secure Shared Access Signature (SAS) tokens.
4. Compute Tier (Azure Functions v4): Executes background image processing, ONNX Runtime SCRFD/ArcFace facial detection, perceptual difference hashing, and OpenAI Foundry GPT-5-mini evaluation.

---

Asynchronous Queue Pipeline

When an analysis job is requested, work items are enqueued to Azure Queue Storage:

Workers scale horizontally to process batches in parallel, updating the database status from INITIATING to PROCESSING and finally COMPLETED.

---

0ms Latency Guarantees

Every rating, tag, and culling decision made by a photographer reflects immediately in the browser via TanStack Query optimistic mutations. The user never waits on HTTP round-trip latency to advance to the next frame. If a network interruption occurs, the client automatically rolls back to the verified server snapshot with an informative toast notice.