Satark-AI

The Problem

Generative AI has made it trivially easy to clone voices, fabricate images, and spread synthetic media. Existing detection tools are siloed — they handle either audio or images, never both — and often require expensive infrastructure or remain offline-only.

The Solution

Architected a scalable Turborepo monorepo with three independent microservices: a React 18 PWA frontend, a Hono (Node.js) API gateway backed by PostgreSQL via Drizzle ORM with Clerk JWT auth, and a Python 3.11 FastAPI AI engine for audio deepfake detection (Wav2Vec2) and speaker biometrics (ECAPA-TDNN). Image deepfake detection runs on a fully serverless Cloudflare Worker proxy that forwards requests to NVIDIA NIM (Llama 3.2-90B Vision), completely independent of the Python engine. A GitHub Actions cron job pings both Render services every 14 minutes to prevent free-tier cold starts.

Key Challenges

• 1Designing a serverless image detection pipeline via Cloudflare Workers with strict CORS whitelisting, 5MB double-layer size enforcement, and 30s AbortController timeout against NVIDIA NIM — all without touching the Python engine.
• 2Implementing lazy model loading and thread executor offloading in FastAPI to prevent OOM crashes and keep the async event loop non-blocking on free-tier Render instances.
• 3Building per-user scoped speaker isolation — users can only verify against their own enrolled voice prints — with 192-dim ECAPA-TDNN embeddings stored in PostgreSQL and cosine similarity computed server-side (threshold: 0.75).

Tech Stack