Edge-AI engineer focused on running computer-vision models on constrained edge hardware — from on-device inference and training-pipeline optimization to a distributed, self-healing and security-hardened edge fleet. Hands-on contributor to a live 100-node leopard-detection network.
Projects
Leopard Detection System — Edge-AI Deployment
Real-time wildlife detection on a distributed NVIDIA Jetson edge fleet for a forest department.
- On-device inference: deployed the CNN detector at 28 FPS per Jetson with TensorRT — all processing at the edge for low-latency, offline-capable detection with no cloud round-trip.
- GPU-load-adaptive frame processing: dynamically tuned input FPS, resolution and batch size from live GPU & thermal telemetry to hold latency in budget on constrained hardware.
- Distributed camera failover: on node failure, re-homed the camera stream to the nearest healthy Jetson in seconds (heartbeats + nearest-neighbour scheduling) — zero coverage gap.
- Optimised model-training pipeline: parallel DataLoader, pinned-memory async transfers and mixed precision (AMP) — cut training time by a large margin and lowered cloud GPU-hours.
- Secured the edge fleet (cyber-hardening): ran every node on a private WireGuard VPN with mTLS and no public inbound ports, plus on-device IP / process whitelisting and file-integrity checks to detect tampering or malware.
- OTA model rollouts across the fleet with staged canaries and automatic rollback.
Containers & Fleet
- Containerised the entire edge stack with Docker for identical runs on laptop, edge and cloud.
- Kubernetes-style self-healing and rolling updates to keep the 100-node fleet current with zero downtime.
- Wrote Bash automation to push commands across all nodes in parallel over SSH.
Advanced AI Ecosystem Training
by DeepManthan AI Production AI Ecosystem Program
Intensive program on building & operating production AI systems — containers, orchestration, distributed systems, networking and deep learning.
Containerisation — Docker, images, isolation, OTA rollouts
Orchestration — Kubernetes, self-healing, scaling
Distributed Systems — Raft consensus, consistent hashing
Messaging — pub/sub, topics, partitions, offsets
Concurrency — multi-threading, producer–consumer
Computer Vision — CNNs, convolution, detection
Networking — TCP/IP, SSH, private overlays
Linux & Ops — processes, daemons, systemd
Impact Highlights
100
edge nodes in the live fleet
28 FPS
on-device inference per Jetson
>90%
GPU utilisation after tuning
99.9%
coverage through node failures