Back to projectsAI RECOVERY

From 12% to 80% Accuracy — A Broken System Made Production-Ready

Took a failing facial recognition deployment from unusable to operational in federal security environments — by solving the problem the original system couldn't see.

Client Type

Federal Agency

Timeline

TBC

Role

Lead AI Engineer

— THE CHALLENGE

A federal security deployment couldn't match faces from live camera feeds to passport photos — and it had to run offline.

The client had deployed a facial recognition system across federal buildings to identify individuals against a passport photo database. The system was failing at 12% accuracy — effectively useless.

Two problems compounded each other: CCTV cameras captured faces at oblique angles and low resolution, while the reference database contained only front-facing passport photos. The angle mismatch made direct comparison nearly impossible. On top of that, the system had to run entirely offline on standard GPU hardware across 20+ simultaneous camera feeds — no cloud, no external APIs, no room for slow inference.

12% → 80%

Recognition accuracy improvement

24 FPS

Near real-time on 20+ simultaneous cameras

Offline

Fully air-gapped, no cloud dependency

— THE APPROACH
01

High-Performance Detection Pipeline

Built a complete facial recognition pipeline from scratch using ArcFace — detection, cropping, normalization, and recognition. Optimized with FAISS-indexed cosine similarity to achieve near real-time processing at 24 frames per second across 20+ concurrent camera streams on standard GPU hardware.

02

GAN-Based Face Synthesis

This was the breakthrough. Developed a GAN model that could take oblique, low-quality CCTV captures and synthesize a front-facing passport-style representation. This solved the fundamental angle mismatch that was destroying the original system’s accuracy.

03

Hardware Optimization

Developed custom CUDA kernels and optimized the pipeline for TensorRT-enabled hardware. The entire system ran offline on air-gapped infrastructure — no cloud calls, no external dependencies, meeting strict federal security requirements.

04

End-to-End Integration

Delivered a production system that could ingest live feeds from 20+ cameras, run face detection and GAN synthesis in real-time, match against the passport database, and surface results to operators — all within federal security and privacy constraints.

— THE RESULTS

BEFORE

  • 12% facial recognition accuracy — functionally useless
  • CCTV angle mismatch against passport photos unsolved
  • System couldn’t keep up with multi-camera feeds in real-time
  • No viable offline solution for air-gapped environments

AFTER

  • 80% recognition accuracy — a 6.7x improvement
  • GAN synthesis bridged the CCTV-to-passport angle gap for the first time
  • Near real-time processing at 24 FPS across 20+ concurrent cameras
  • Fully offline, air-gapped deployment meeting federal security requirements

Have a similar challenge?

I help teams design, build, and ship AI systems that work in production. Let's talk about your problem.