Edge-AI systems engineer trained to design and operate production-grade AI ecosystems end to end — from on-device computer-vision inference to distributed, self-healing, observable deployments. Hands-on contributor to a live 100-node leopard-detection network for a forest department, shipping distributed-systems, multi-agent, and adaptive-inference features under real field constraints. Also optimises production model-training pipelines for speed and cost, and debugs failures down to the packet level.

Projects

Leopard Detection System — Production Edge-AI Ecosystem
Live deployment · 100 Jetson nodes
Real-time wildlife detection & human–animal conflict mitigation across a distributed camera fleet (NVIDIA Jetson · Kubernetes · private VPN mesh).
  • Distributed camera failover (high availability): engineered automatic re-homing of a failed node's camera stream to the nearest healthy Jetson — heartbeat health-checks detect the outage and a nearest-neighbour scheduler re-binds the RTSP stream in seconds, sustaining detection coverage with zero manual intervention and no single point of failure.
    heartbeatsservice discoveryfailover / HAWireGuard mesh
  • Leopard tracking agent (Multi-Agent System): built an autonomous tracking agent that consumes detections off a message bus and maintains per-animal track IDs across frames and cameras using motion prediction and re-identification, coordinating with perception, alerting, and ranger-dispatch agents to turn raw detections into actionable movement paths.
    multi-agent (MAS)object tracking / Re-IDpub/sub messagingagent coordination
  • GPU-load-adaptive frame processing: implemented a controller that dynamically tunes input FPS, resolution, and batch size from real-time GPU utilisation and thermal telemetry, applying backpressure to hold inference latency within budget and eliminate frame drops during load spikes — maximising throughput per watt on constrained edge hardware.
    adaptive inferencebackpressureGPU schedulingTensorRT
Optimised AI Model Training Pipeline — Production AI Server
On-prem GPU server · PyTorch
Trained the leopard-detection model on a production-grade GPU server and re-engineered the data-to-GPU pipeline to remove idle time and cut training turnaround.
  • Profiled the pipeline end to end — GPU utilisation, VRAM, PCIe bandwidth and dataloader CPU — and pinpointed the classic GPU-starvation bottleneck, where the GPU sat idle waiting on disk I/O and CPU preprocessing.
  • Eliminated data-loading stalls: tuned parallel DataLoader workers for concurrent disk reads, enabled pinned_memory + non_blocking async PCIe transfers, and prefetched batches so the CPU keeps loading while the GPU trains — lifting GPU utilisation from ~50% to >90%.
  • Accelerated compute: applied mixed-precision (AMP / FP16) training, right-sized batch size to VRAM, and streamlined augmentation — cutting epoch and total training time by a large margin with no drop in accuracy.
  • Cost impact: shorter training turnaround on-prem translates directly into major cloud-cost savings for teams training on rented GPU instances — fewer GPU-hours per model.
pipeline profilingDataLoader tuningpinned / async transfermixed precisioncloud-cost savings
Edge & AI Server Threat-Monitoring Agent — Cybersecurity for the Fleet
Python · Linux · runtime security
Wrote a lightweight host- & network-monitoring script running on AI server & every edge node to detect suspicious activity and enforce a locked-down, zero-trust security posture.
  • IP allow-listing: restricted each node to communicate only with a known set of trusted endpoints (cloud, peer nodes, VPN), flagging and blocking any connection to an unlisted address.
  • Process allow-listing: continuously audited running processes against an approved baseline, alerting on any unexpected or unsigned binary to catch tampering or malware.
  • Outbound socket monitoring: inspected all outbound socket connections for anomalies — unknown destinations/ports, beaconing patterns, and potential data-exfiltration attempts.
  • Disk-activity monitoring: watched file-system changes on sensitive paths (models, configs, keys) to detect unauthorised writes and ransomware-style bulk modification.
  • Alerting: streamed severity-tagged security events to the central dashboard, feeding the ops/ranger alerting pipeline for rapid response.
IP allow-listingprocess allow-listingsocket monitoringfile-integrity (FIM)threat detection
Edge & AI Observability Dashboards — Single Pane of Glass
Real-time · 100-node fleet
Designed and built rich, real-time dashboards that give operators one view over both the infrastructure health and the AI model's behaviour across the fleet.
  • Edge Observability: streamed live per-node telemetry — CPU, GPU, memory, temperature, disk & network — into a rich operations dashboard (fleet KPIs, active / degraded / offline counts, and per-device resource gauges), giving early warning of hardware stress before a node fails.
    real-time telemetryresource gaugesfleet KPIsrich dashboard UI
  • AI Observability: tracked model-quality metrics in production — accuracy / precision / recall, false positives & negatives, per-class confidence and detection drift — alongside inference latency (p50 / p95), throughput (FPS) and GPU-time per frame, surfacing model regressions the moment they appear.
    model metricsfalse-positive rateinference latencydrift detection
  • Actionable alerting: set thresholds and SLOs (FP-rate spike, latency breach, node degraded) wired into the ops pipeline — turning raw dashboards into automatic, prioritised alerts.

Debugging & Diagnostics

Fleet Automation & Shell Scripting

Advanced AI Ecosystem Training

by DeepManthan AI Production AI Ecosystem Program
Intensive program on building & operating production AI systems — from containers and orchestration to distributed systems, networking, and deep learning.
Containerisation — Docker, images, isolation, OTA rollouts
Orchestration — Kubernetes, self-healing, scaling, rollouts
Distributed Systems — Raft consensus, consistent hashing
Messaging — pub/sub, topics, partitions, offsets & lag
Concurrency — multi-threading, producer–consumer
Computer Vision — CNNs, convolution, object detection
Networking — TCP/IP stack, SSH, private overlays
Linux & Ops — processes, daemons, systemd, observability
Data & APIs — SQL, REST, HTTP, JSON
DS & Algorithms — lists, hash tables, trees, complexity

Impact Highlights

100
edge nodes in the live production fleet
99.9%
detection coverage maintained through node failures
28 FPS
real-time on-device inference per Jetson
>90%
GPU utilisation after pipeline tuning (from ~50%)
↓ Training
large cut in training time → lower cloud GPU-hours
Sample resume...