Owl Vision - Pranav Bonthu

Project Overview

At Owl Vision I built the machine learning side of a pest-monitoring product for farms. Camera traps in the field photograph insects, and a YOLO11m computer vision model finds and classifies the pests in each image. Farmers see which pests are showing up in real time, so they can pick the right pesticide and avoid spraying chemicals they don't need.

I also worked as a field engineer: I drove out to farms to install, troubleshoot, and maintain the camera hardware.

Key Features

Pest Detection

YOLO11m object detection model that finds and classifies invasive pest species in trap images

GPU Training Pipeline

Dockerized training pipeline on GCP GPU VMs that cut YOLO11m training time by 4x

Field Hardware

Cellular, GPS-synced camera devices deployed on farms for continuous monitoring

Farmer Impact

Real-time pest identification helps farmers buy the right pesticides and cut unnecessary chemical use

ML Training Pipeline

Training used to be slow and hard to reproduce, so I packaged it into a containerized pipeline that runs the same way on any GCP GPU machine.

Pipeline Highlights:

Trained on a dataset of 9,460 labeled image tiles pulled from Google Cloud Storage
Full training runs of 150 epochs on NVIDIA GPU VMs
Docker image bundles CUDA, drivers, and dependencies so any fresh VM can train with a single command
Scripted setup: provision the VM, install Docker and the NVIDIA driver, ship the code, build, train, then delete the VM to control cost
Smoke-test mode for checking the pipeline end to end before starting a multi-hour run

Field Deployment

In the Field:

First farm test in July 2026 with networked camera devices
Devices report firmware, GPS time sync, and cellular signal health for remote diagnostics
On-site troubleshooting and maintenance of hardware at partner farms

Beehive Project (Feb–May 2026)

Before the internship, I worked on the Olin Bee Team, a five-person Olin x Owl Vision project to help beekeepers monitor colony health and catch Varroa mites early. I was on the software side of the team.

We interviewed beekeepers, from hobbyists to a commercial operation with over a thousand hives, to learn their biggest challenges. We looked into thermal mite detection and ruled it out: the research was contradictory, and affordable thermal cameras don't have the resolution to see the ~0.65°C difference. We ended up with a handheld vision device with temperature and humidity sensing, aimed at sideliner and hobbyist beekeepers.