Precision AgTech on
$100 Phones
Real-time pest detection, biological substitution logic, and drone spot-spraying maps running 100% offline on low-end ARM64 silicon. Slashing pesticide volume up to 80% for smallholder farms. No internet. No heavy machinery.
Architecture
Three Pillars of Edge Precision Agriculture
Real-Time Edge Vision
The Pocket Sensor
Replacing $400K tractor sensor rigs with the phone camera. Runs int8-quantized YOLOv11-Nano natively via NDK/C++ in <30ms. Identifies pests (Mustard Aphids, Fall Armyworm) and early leaf blights before farmers can notice or describe them.
Biological Substitution Engine
Bio-First Rules
Hard-coded Economic Threshold Levels (ETL). If pest pressure is early, synthetic chemicals are locked out — system prescribes ICAR-verified biologicals (Trichoderma, Beauveria bassiana, Neem oil) with physics-aware constraints: late-afternoon spray windows to prevent UV microbial death.
Precision Drone Mapping
Spot-Spray Integration
As scouts walk field transects, infected zones are GPS-tagged offline with sub-meter relative accuracy. The app outputs an instant KML/GeoJSON flight path compatible with Indian Kisan Drones (IoTechWorld, Garuda), triggering nozzles only over infected crops. Turns blanket spraying into micro-targeted interventions.
On-Device Pipeline
Multimodal Edge Architecture
Every step runs locally on the phone. No cloud calls. No data leaves the device.
OR Hindi Voice → Shruti STT (29MB)
Dual modality, single device
Below ETL → bio-only lock
Above ETL → chemical authorized
BM25-ranked advisories
Dosage: locked to verified registry
5KB KML drone mission file
Knapsack spot-spray instructions
Edge Runtime
Runtime Provenance
The underlying C++/NDK execution harness was architected for high-concurrency, memory-constrained mobile environments, sustaining zero-crash inference over 10,000+ benchmark cycles on 2GB RAM devices.
Founder
Built by One Engineer
Spent a decade building ultra-low-latency distributed systems at Facebook and 100ms. Now applying edge systems engineering and embedded model optimization to solve smallholder agricultural chemical dependency where the cloud cannot reach. IIT Delhi CS. Codeforces Grandmaster (2416, 3rd in India).
Blog
Engineering Notes
Compiling YOLOv11-Nano and llama.cpp on Android NDK for Sub-30ms Field Inference
Cross-compiling vision and language models for ARM64 with CMake, JNI bridge, and proper UTF-8 Devanagari handling on $100 phones.
Deterministic ETL Safety Gates: Restricting Agrochemicals via Embedded SQLite FTS5
How hard-coded Economic Threshold Levels and BM25-ranked ICAR advisories enforce bio-first prescriptions on-device.
Deploy in the Field
Deploying offline precision agronomy with FPOs, drone service hubs (Drone Didis), and agricultural co-ops across North India. Looking for agronomists, edge ML engineers, and YC partners.