Edmund Jeevan D'Souza

Edge Vision Actuation — Voice-Controlled Autonomous Robot

A full perception-navigation-task planning integration on real hardware: an NVIDIA Jetson Orin Nano Super mounted on a mobile robot chassis, closing the loop from camera and microphone input to real motor control — no cloud dependency, all inference on-device.

Exported a YOLOv8 detector to TensorRT FP16, benchmarking a measured 27% throughput improvement (13.9 → 17.6 FPS) over baseline PyTorch inference. Built a GPU-accelerated voice pipeline (wake-word detection + Whisper transcription) driving the robot's motors over a serial Open Interface protocol, with live obstacle-safety behaviour confirmed on hardware. Diagnosed and fixed a safety-critical integration bug where obstacle-detection events silently disabled further commands.

Next: Camera-Based Obstacle Avoidance & Pan-Tilt Tracking

In progress: merging the YOLOv8 detection pipeline with the voice-control loop for autonomous obstacle avoidance, plus closed-loop pan-tilt camera tracking. Stereo depth from the second IMX219 sensor and a LeRobot manipulation policy on a low-cost arm remain on the longer-term roadmap.

REGLO — Renewable Energy Grid Load Optimiser · In development

A duck-curve early-warning service for the UK grid: an hourly pipeline pulls Copernicus satellite data (CAMS irradiance, ERA5 wind, C3S temperature), computes per-zone net load, scores belly-depth, ramp and curtailment risk, and dispatches severity-graded alerts by email, SMS or webhook into SCADA systems — up to 6 hours ahead of the event window.

Designed end-to-end around UK DNO geography and the NESO demand baseline, with an audit trail comparing predicted metrics against actuals. Built on free, open European satellite infrastructure.

Lithium-Ion Battery Element Analysis — ML POC

Independently architected and delivered a five-stage computer-vision pipeline analysing elemental composition of Li-ion battery cross-sections from 120-slice TIFF sequences: dual-pass Ilastik segmentation, custom annotation, k-means colour clustering (OpenCV, scikit-learn) and a Mask RCNN model trained on local NVIDIA GPU hardware.

Demonstrated open-source tooling matching proprietary prediction accuracy for a production engineering client. Formerly under NDA; now publicly disclosable.

Hackathon Builds — Winner, UK Runner-Up & Event Ops

Winner at TECHXLR8 London (ExCeL, 2018, hosted by Digi.me): working prototype integrating personal-data APIs inside a 24-hour sprint — subsequently co-founded Health Haven Technologies to take the winning concept to market. Runner-up and Best App in the UK at Global DATAHACK14. 21 hackathons entered to date — a standing habit of shipping working prototypes under deadline pressure.

Also on the other side of the table: part of the organising team for the AI Voice & Meeting Agents Hackathon in London, run by meetstream.ai with Antler UK — a 76-attendee, full-day event at Techspace Aldgate East. Ran event logistics, the registration desk and participant welcome, and provided on-site tech support to keep 20+ teams building all day.