← Back to work
TMA Solutions
Threat Detection Device
A phone app that spots threatening objects, speech and behaviour in real time, fully on-device. Running in production for the customer.
- Role
- AI Engineer | AI Researcher
- Context
- TMA Solutions
- Period
- Jan 2025 – Oct 2025
Overview
A cross-platform mobile protection application that identifies threatening behavior in real time from video and audio streams and escalates alerts to the appropriate endpoint using a set of deep learning models.
Key achievements
Deployed to production with real-world usage on the customer side.
Delivered a golden dataset and a set of AI models that met 90% of the customer's required testing criteria.
Responsibilities
- Researched and surveyed mobile-deployable models for computer vision and NLP tasks.
- Collaborated with customers to collect and procure data for threat object detection, as well as threatening text and voice samples.
- Fine-tuned a YOLO model for object detection and deployed it on Android with TensorFlow; implemented custom YOLO post-processing on iOS to improve on the built-in framework's processing time.
- Built an audio processing pipeline using OS built-in speech-to-text, MobileBERT for NLP, wav2vec, and a custom RCNN-based audio model.
- Fine-tuned MobileBERT and the custom RCNN-based model on the collected dataset and deployed them with Google MediaPipe on Android and iOS.
- Built a CNN-based facial emotion classification model.
- Developed a full multi-model pipeline for real-time processing while reducing resource usage on edge devices.
Tech stack
Languages & Frameworks
PythonTensorFlowGoogle MediaPipe
Models
YOLOMobileBERTwav2vecCNNRCNN
Platforms
AndroidiOS