When you build machine learning systems in an air-conditioned laboratory in Bengaluru, the cloud feels infinite. You have gigabit fiber, elastic AWS or GCP clusters, and millisecond access to the world’s largest foundational models.
Then you take your hardware into rural Chitradurga, stand inside a dairy farming shed where cellular reception drops to absolute zero, and realize: if your AI cannot function without the internet, it is essentially useless.
In October 2026, I joined Novabion Pvt. Ltd. as a Project Engineering Intern under the leadership of Managing Director Girimallikarjuna SMD. My mission was explicit and tactical: execute the Phase 1 MVP of Pashu Drishti — an offline-first Edge AI system engineered to bring real-time livestock health monitoring and diagnostics straight to grassroots agricultural environments.
This is the story of how we architected, optimized, and physically deployed Pashu Drishti in the field.
The Core Challenge: Zero-Cloud Livestock Intelligence
Livestock management is the economic lifeline of rural Karnataka. For smallholder farmers, a single diseased cow or late-detected complication can wipe out months of household income. While urban dairy conglomerates can afford connected IoT arrays, rural farmers face stark realities:
- No Cellular Connectivity: Barns and grazing fields are frequently in cellular dead zones.
- Extreme Physical Constraints: Dust, high ambient temperatures, fluctuating local power grids, and unpredictable animal movement.
- Latency Sensitivity: An animal diagnostic assessment needs to occur in real time as the animal is in the corral or examination chute, not queued for an asynchronous cloud batch job hours later.
We needed a system that runs 100% locally on edge compute hardware, captures visual and thermal telemetry, runs deep learning inference directly on the device, and stores verified records for opportunistic synchronization whenever a network connection is detected.
Architectural Blueprint: The Pashu Drishti Edge Pipeline
Pashu Drishti's edge architecture is separated into three decoupled, high-resilience layers:
+-------------------------------------------------------------+
| PASHU DRISHTI EDGE STACK |
+-------------------------------------------------------------+
| |
| [ Optical & Thermal Sensor Feeds ] |
| │ |
| ▼ |
| [ Ingestion & Frame Preprocessing ] (OpenCV / Zero-Copy) |
| │ |
| ▼ |
| [ Local On-Device Neural Engine ] |
| ├── Quantized Feature Extractor (INT8) |
| ├── Anomaly & Disease Detection Classifier |
| └── Rule-Based Diagnostic Validation Guardrail |
| │ |
| ▼ |
| [ Embedded Edge Database ] (Local SQLite / WAL mode) |
| │ |
| ▼ |
| [ Opportunistic Sync Daemon ] |
| └── Syncs securely to central cloud when connected |
| |
+-------------------------------------------------------------+
1. Zero-Copy Sensor Ingestion
Livestock are rarely cooperative subjects. Video feeds and camera frames suffer from motion blur, sudden occlusions, and severe lighting variation. We implemented a lightweight preprocessing pipeline in C++ and Python using optimized OpenCV bindings:
- Real-time frame quality scoring (discarding blurred frames before they waste inference compute).
- Adaptive histogram equalization to normalize harsh sunlight and deep barn shadows.
- Zero-copy memory buffers passing raw tensors straight to the edge runtime without duplicate memory allocations.
2. Quantized Local Inference Engine
Running unquantized 32-bit floating-point neural networks on resource-constrained edge hardware causes severe thermal throttling and battery drain.
We quantized our vision models down to INT8 precision using post-training calibration datasets reflecting real farm imagery:
# Edge Inference Runner with Quantized Runtime
import onnxruntime as ort
import numpy as np
class PashuDrishtiEdgeEngine:
def __init__(self, model_path: str):
# Configure hardware execution provider (NPU / GPU / CPU)
opts = ort.SessionOptions()
opts.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
opts.intra_op_num_threads = 4
self.session = ort.InferenceSession(
model_path,
sess_options=opts,
providers=['CPUExecutionProvider']
)
self.input_name = self.session.get_inputs()[0].name
def evaluate_livestock_frame(self, preprocessed_frame: np.ndarray) -> dict:
# Zero-latency local execution - under 45ms per frame
raw_outputs = self.session.run(None, {self.input_name: preprocessed_frame})
confidence, diagnosis_code = self.parse_telemetry(raw_outputs)
return {
"status": "HEALTHY" if confidence > 0.85 else "ANOMALY_DETECTED",
"confidence": float(confidence),
"code": diagnosis_code,
"processed_locally": True
}
By optimizing memory locality and using hardware-accelerated vector instructions, we sustained inference latency under 45 milliseconds per frame on compact, fanless hardware.
3. Local-First Embedded Persistence
Instead of failing when HTTP requests time out, all diagnostic results, timestamps, and model logs are written immediately to a local embedded database operating in Write-Ahead Logging (WAL) mode.
An opportunistic background daemon monitors network interfaces. When the device returns to a base station or detects Wi-Fi/cellular connectivity, it performs an authenticated, atomic sync transaction to Novabion's central analytics backend.
Field Deployment in Chitradurga: Where Theory Meets Dirt
The real validation of Pashu Drishti didn't happen in a notebook — it happened on the ground in Chitradurga.
Deploying AI in rural agricultural conditions tests variables that academic papers never discuss:
- Dust and Moisture Ingress: Farm environments are punishing. Enclosures had to be sealed and passively cooled because active cooling fans quickly choke on particulate matter.
- Dynamic Animal Stance: Unlike humans who can be instructed to hold still for a scan, cattle shift weight, turn away, and interact with the herd. We had to implement multi-frame temporal voting so that momentary bad angles didn't trigger false positives.
- Veterinary Usability: A user interface for rural field workers must be uncluttered, high-contrast (readable under direct sunlight), and provide immediate color-coded alerts (Green: Normal, Amber: Monitor, Red: Immediate Veterinary Attention).
Seeing the physical system successfully identify early diagnostic indicators on real animals in Chitradurga — with zero internet connection and instant feedback — was one of the most rewarding engineering milestones of my career.
Key Engineering Lessons from the Novabion Sprint
- Edge AI is a Systems Discipline, Not Just a Modeling Discipline: The best model architecture is useless if your memory footprint exceeds device RAM or if thermal throttling cuts your clock speed in half after 10 minutes of continuous operation.
- Offline-First is a Core Requirement: Building software for developing markets demands designing for the worst-case connectivity environment from Day 1.
- Milestone Velocity Matters: In a 30-day tactical execution sprint, there is no time for architectural bloat. You write clean, modular code, establish strict testing benchmarks, and iterate rapidly based on real physical deployments.
Working with Novabion Pvt. Ltd. proved that modern AI can step out of the cloud and solve foundational agricultural problems right where it matters most.
Continue Reading
Curated deep-dives in AI systems and engineering architecture.
Deploying GGUF Models for On-Device Inference
A practical guide to packaging compact GGUF models for local inference with llama.cpp, optimizing for zero-latency, absolute privacy, and offline edge constraints.
Inside AatonovaZ: Architecting Production Generative AI Systems and Enterprise LLM Pipelines
Reflections on my Gen AI Engineering internship at AatonovaZ Technologies — engineering multi-stage LLM data pipelines, prompt evaluation benchmarks, and hardware-software AI integration directly with executive leadership.