Koca Ventures Ltd
71-75 Shelton Street
Covent Garden, London
WC2H 9JQ, United Kingdom
Registered in England & Wales — 16231043
Portfolio
Production systems we designed, built, and run — agentic automation, internal tools, computer vision, and B2B platforms. Client names stay private.
Project Orion Restaurant AI
AI-powered restaurant operations platform that turns existing security cameras into intelligent service monitors. Built for a hospitality group whose platform spans 150+ locations, with computer vision live at 10+ sites, the system detects customer attention signals, tracks table status, and routes voice commands to staff in real time — integrating with existing POS systems with no additional hardware.
Project Vanguard Security
Multi-camera perimeter security system built on NVIDIA DeepStream, ingesting up to 16 IP camera streams through GStreamer pipelines on Jetson Orin Nano edge nodes. Custom-trained ReID and behavior-analysis models detect intrusion, loitering, and tailgating in real time, with TensorRT FP16 optimization engineered to a sub-20 ms per-stream latency budget. An MQTT event bus pushes alerts to a React command dashboard, validated through a staged pilot before full-site production rollout.
Warehouse Vision System
Edge-deployed computer vision for package tracking, inventory counting, and route optimization across a warehouse floor. Jetson AGX Orin nodes run DeepStream pipelines processing ceiling-mounted IP cameras, while ESP32-S3 microcontrollers on custom PCBs handle barcode-trigger sensors and conveyor-gate signals over MQTT. We collected and annotated the package-image dataset on-site, then used TensorRT pruning and INT8 quantization to reach real-time throughput on every node.
Predictive Maintenance AI
IoT sensor network using custom-designed PCBs built around ESP32-S3 modules, capturing vibration, temperature, current-draw, and acoustic data from CNC machines on the shop floor. Sensor firmware streams data over MQTT to a TimescaleDB time-series store, where lightweight TensorFlow Lite models run anomaly detection at the edge — built to flag developing bearing and spindle failures before they stop the line.
Fleet Management Platform
GPS and telemetry platform for a logistics fleet, powered by custom PCBs with ESP32 modules, u-blox GNSS receivers, and CAN-bus adapters. Each unit streams position, fuel level, and engine diagnostics over LTE-M to a Node.js ingest service backed by PostgreSQL with PostGIS. The React dashboard provides live map tracking, geofence alerts, route replay, and driver-behavior scoring for fuel and delivery-window management.
Facial Recognition Access
Facial recognition access control deployed on Jetson Orin Nano edge devices at building entry points, using ArcFace embeddings optimized through TensorRT FP16. GStreamer pipelines handle camera ingestion with infrared and visible-light dual-sensor fusion for anti-spoofing. Employee enrollment runs through a secure on-premise pipeline with no cloud dependency, tuned for a strict false-accept threshold across varying lighting conditions.
Production Line Analytics
Real-time OEE and cycle-time analytics dashboard ingesting multiple GStreamer camera feeds processed on Jetson AGX Orin via DeepStream, combined with PLC signals from the shop floor. A Next.js frontend with D3.js visualizations and TimescaleDB continuous aggregates delivers per-station efficiency metrics, bottleneck detection, and shift-comparison reports — surfacing micro-stoppages invisible to manual observation.
Smart Warehouse Robotics
Autonomous mobile robot navigation system combining Jetson AGX Orin vision processing with ROS2 and LiDAR-based SLAM for warehouse operations. DeepStream pipelines on the robots detect pallets, obstacles, and human workers in real time, while ESP32-based custom PCBs bridge the robot control bus with warehouse management signals over MQTT. We built the synthetic training dataset using domain randomization in NVIDIA Isaac Sim, then transferred the models to physical robots with a limited set of real-world fine-tuning data.
Retail Video Analytics
Customer flow and heatmap analytics for a multi-store retail chain, running NVIDIA DeepStream on Jetson Orin Nano nodes processing the IP cameras at each location. TensorRT-optimized person detection and tracking models generate dwell-time, queue-length, and zone-transition metrics stored in PostgreSQL. A React analytics dashboard lets store managers compare foot-traffic patterns across locations and time periods to inform layout decisions.
Ready to Start Your Project?
Tell us what you need — we'll scope it honestly and show you what we'd build on infrastructure you own.
Get in Touch