Note on Who Should Not Apply

This role is NOT suitable for candidates whose experience is primarily in LLMs, generative AI, prompt engineering, RAG, agentic workflows, or API integration. Candidates limited to YOLO fine-tuning, OpenCV pipelines, video analytics, pretrained models, copied repositories, or notebook demonstrations should also not apply. We require deep, hands-on ownership of CNNs, Vision Transformers, dense-prediction architectures, model training, failure diagnosis, evaluation, and production engineering.

Employment Type: Full Time || Remote(Preference for candidates in NCR)

About what you will do

You will own difficult multiclass semantic segmentation problems on high-resolution aerial imagery across architecture, data, training, evaluation, tooling, and production.
You will diagnose why individual classes and boundaries fail, prove root causes through controlled experiments, and fix them through architecture, loss, supervision, optimisation, or data - not cosmetic postprocessing.
You will build and modify CNNs, Vision Transformers, and hybrid systems at the architecture and training-loop level; merely fine-tuning existing repositories is not sufficient.
You will train and optimise large models on A100/H100-class GPUs, work through imperfect datasets and codebases, and build practical internal applications when required.

What you need to have

3-5 years of serious hands-on experience with CNNs, Vision Transformers, multi-scale features, dense prediction, loss functions, and class-wise evaluation. Strong professional experience with semantic or instance segmentation is expected.
Strong Python and PyTorch skills, including custom training, custom loss functions and evaluation pipelines.
Experience owning GPU training runs and diagnosing memory, throughput, convergence, and distributed-training failures.
Strong data-engineering and software-engineering fundamentals.
Experience with Linux, Git, Docker, SQL, NumPy, Pandas or Polars, testing, logging, and experiment tracking.
Ability to clearly explain unsuccessful experiments, failure analysis, technical trade-offs, and measurable improvements.
Willingness to perform both architecture-level work and unglamorous implementation, debugging, and data-cleaning work.

What is good for you to have

Experience with aerial, satellite, geospatial, medical, microscopy, or other high-resolution segmentation domains.
Experience with geospatial rasters, tiling, overlapping inference, GeoTIFFs, GDAL, or Rasterio.
Relevant master's or PhD degree; not required.

About SiteRecon:

SiteRecon is on a mission to revolutionize the way businesses make decisions in North America by helping enterprise customers scale their sales teams. We automate property mapping and site visits through our innovative mapping platform, designed specifically for property maintenance contractors. By streamlining these processes, we give our clients back valuable time, accelerate their sales efforts, and enhance their profitability. Join us in building the next-generation property intelligence platform!

To learn more about what we do for our customers - https://order.siterecon.ai

About the Product:

SiteRecon combines the power of Amazon and Google Docs for maps, allowing users to easily create customized maps through our AI-driven data delivery system. This platform serves as a foundation for creating survey notes, operational plans, cost estimates, and more.

Market Overview:

Our primary focus is the landscaping industry in the United States, Canada, and Australia, which collectively represents a staggering $200 billion market.

Deep Learning Engineer Vision & Geospatial AI