Senior Computer Vision Researcher

Following selection criteria will be followed -

5-15 YOE

Worked on at least 1 Computer vision problem statement in the past 5 years

Have built, trained and deployed models on production before

Comfortable with WFO in Bangalore or Noida

Happy to read research papers day in day out and train on terabytes of image data on big H100 clusters

preferably tier-1 college ( IISc, IITs, NITs, BITS and others)

About Us

SiteRecon is a B2B SaaS platform transforming property measurements for landscapers and snow removal contractors in the $176B US market. Using advanced AI on high-resolution aerial imagery, we’ve reduced measurement time from 1 week to 1 day, aiming for 1 hour with upcoming AI advancements.

Role Overview

We seek a skilled engineer to evaluate, enhance and / or transform our computer vision infrastructure from traditional CNN architectures to cutting-edge transformer-based models. This role demands expertise in model design, training, and optimization.

You’ll have access to :

A vast aerial imagery database (7 cm GSD) of 500,000+ properties across the U.S. since 2021.

A team of 60 annotators mapping thousands of acres daily.

A ready market of 350+ customers for immediate deployment.

Powerful compute resources for rapid model training.

This is a frontier challenge in computer vision and GIS. While solutions like Meta’s SAM offer basic raster segmentation, they lack the precision for creating high-fidelity, topologically consistent vectors essential for practical GIS applications. SAM struggles with occlusions (e.g., shadows, tree canopies) and produces "blobs" rather than architect-quality outputs.

Our approach focuses on solving a constrained problem : using the world’s highest-resolution aerial imagery (7 cm) over U.S. urban areas with logical, repeatable patterns. By tackling this focused challenge, we aim to develop scalable templates to generalize automated extraction for broader GIS applications, similar to Waymo’s strategy in self-driving technology.

Key Responsibilities

Design and implement transformer-based architecture for semantic segmentation of aerial imagery

Develop efficient image-to-token and token-to-image conversion pipelines

Create and maintain training datasets, including data cleaning, augmentation, and validation

Optimize model training processes for distributed computing environments

Implement efficient inference pipelines for production deployment

Collaborate with engineering team to integrate new models into existing infrastructure

Required Technical Skills

Strong foundation in computer vision and deep learning fundamentals

Extensive experience training transformer models from scratch

Expert-level proficiency in PyTorch

Experience with ONNX or TensorRT model optimization and deployment

Deep understanding of distributed computing and parallel processing

Advanced Python knowledge, including multi-threading and multi-processing optimization

Experience with semantic segmentation tasks

Proven track record of handling large-scale data processing

Required Experience

5+ years of hands-on deep learning experience

Track record of successfully deploying computer vision models in production

Experience with vision transformer architectures

Experience optimizing models for production using ONNX / TensorRT

Background in handling high-resolution satellite / aerial imagery preferred

Masters / PhD in Computer Science, Machine Learning, or related field preferred

Desired Qualities

Strong mathematical foundation in deep learning concepts

Experience with model architecture design and optimization

Proven ability to conduct independent research and stay current with latest developments

Excellence in technical documentation and communication

Self-motivated with a passion for solving complex technical challenges

What Sets You Apart

Experience with vision transformers specifically for segmentation tasks

Published research or contributions to open-source computer vision projects

Experience with high-performance computing environments

Background in geospatial data processing

Hands-on experience with model quantization and optimization using ONNX / TensorRT

Experience deploying optimized models in production environments

Why This Role Matters

Every day without improved segmentation costs us real business opportunities. We need someone who moves fast, thinks systematically, and delivers production-ready improvements quickly.