AI/ML Integration
Production face-recognition and computer-vision features — the kind that run reliably on real users and real hardware, not a Jupyter notebook demo.
What this covers
Most "AI integration" pitches stop at a proof of concept. This is the harder part: getting a face-recognition or computer-vision model to run accurately and fast, on-device or via API, at real-world scale with real edge cases — different lighting, angles, and hardware.
- Face recognition: DeepFace (Facenet512) with cosine-similarity matching for identity verification and attendance.
- Face detection: MediaPipe BlazeFace and ML Kit for fast, on-device detection before matching runs.
- Backend serving: FastAPI/Flask services to run inference reliably at scale.
- Mobile integration: CameraX-based capture pipelines feeding directly into the recognition pipeline on Android.
Representative work
Niharika Attendance System
Live on Play StoreFace-recognition attendance app built on a DeepFace (Facenet512) + MediaPipe BlazeFace pipeline with cosine-similarity matching, served through a FastAPI backend. Running at 99.65% accuracy in production with 7,000+ active users — it replaced a client's paid attendance software entirely, cutting their cost from ₹3,290/month to ₹0.
How it typically goes
Validate accuracy on your real data
Before committing to an approach, the model gets tested against your actual conditions — lighting, hardware, user volume.
Build the serving pipeline
A production FastAPI/Flask service, or on-device inference where latency or offline use matters more.
Integrate and monitor
Wired into your existing app, with accuracy and failure-rate visibility after launch — not a black box.
Need AI that actually works in production?
Tell me what you're trying to detect, recognize, or automate — I'll tell you if it's a good fit and what accuracy to expect.