Projects that moved the needle
Real AI systems we've designed, built, and deployed — from pose estimation for clinics to document intelligence for fintech.
Real-Time Person Re-Identification System
Built a multi-camera person re-identification system for a retail chain. The model matches individuals across non-overlapping camera feeds in real time, enabling foot-traffic analytics and loss prevention without facial recognition.
LLM-Powered Legal Contract Analyser
Developed a retrieval-augmented generation pipeline that reviews legal contracts, flags risky clauses, and generates plain-language summaries. Reduced manual review time by 80% for a mid-size law firm.
Gym Pose Coach — Real-Time Form Correction
A browser-based AI coach that uses pose estimation to detect exercise form in real time, providing instant audio and visual feedback for squats, deadlifts, and push-ups.
Automated Crop Disease Detection POC
Proof-of-concept mobile app that identifies crop diseases from leaf photos. Trained on 50k+ images covering 38 disease classes across 14 crop species with 96% validation accuracy.
Voice AI Appointment Booking Agent
An AI voice agent that handles inbound calls, understands patient requests via NLU, checks calendar availability, and books appointments — all in natural conversation with <2s latency.
Surgical Instrument Detection for OR Analytics
Real-time detection and tracking of 23 surgical instrument classes during laparoscopic procedures. Deployed on edge devices in the operating room for workflow analysis and safety compliance.
INT8 Quantization of DETR for Edge Deployment
Quantized a DETR object detection model from FP32 to INT8 using quantization-aware training, achieving 3.8× speedup on NVIDIA Jetson with less than 1% mAP drop.
Customer Churn Prediction ML Pipeline
End-to-end ML pipeline predicting customer churn for a SaaS platform. Features automated retraining, drift detection, and Slack alerts — integrated with the client's CRM for proactive retention.
Multi-Speaker STT with Diarization
Real-time speech-to-text system with speaker diarisation for meeting transcription. Handles up to 8 concurrent speakers with per-speaker timestamps and confidence scores.
Human Action Recognition for Warehouse Safety
Video-based action recognition system for warehouse safety monitoring. Detects unsafe behaviors like improper lifting, missing PPE, and restricted zone entry — triggering real-time alerts.
Document Intelligence for Invoice Processing
Automated invoice processing pipeline combining OCR, layout understanding, and LLM extraction. Handles 15+ invoice formats, extracts line items, and pushes structured data to the client's ERP.
Athlete Sprint & Biomechanics Analysis POC
Proof-of-concept 3D biomechanics platform for sprint analysis. Reconstructs full-body 3D pose from monocular video, calculates joint angles, stride length, and ground contact time.
P&ID Engineering Drawing Automation
Automated parsing of piping and instrumentation diagrams (P&IDs). Detects symbols, reads text labels, and reconstructs the process graph — replacing weeks of manual digitisation.
ADAS Pedestrian Pose Estimation
Pedestrian pose estimation module for an ADAS system. Predicts body pose and intent (crossing, waiting, walking away) to improve autonomous braking decisions at intersections.
Hallucination-Free Legal Contract RAG
A retrieval-augmented generation system for legal Q&A with built-in hallucination detection. Every answer is grounded in source clauses with confidence scores and citation links.
E-Commerce Virtual Try-On & Photoshoot AI
AI-powered virtual try-on and product photoshoot platform for e-commerce. Generates model-on-garment images and lifestyle product shots — eliminating the need for physical photo shoots.
Manufacturing Weld Defect Inspection
Automated weld defect detection system using thermal and visible-light cameras. Classifies 7 defect types in real time on the production line with sub-second inference on edge hardware.
AI Recruitment & Candidate Matching
ML-powered recruitment platform that parses CVs, extracts skills, and ranks candidates against job descriptions using semantic similarity — cutting screening time from days to minutes.