Computer Vision, Applied Research Scientist

Boon Technologies, Inc.•San Francisco, CA

About The Position

We are building the first foundation model for construction drawings — a unified multi-modal vision system that reads, understands, and reasons about architectural, mechanical, electrical, plumbing, and structural plans the way a human estimator does. As a Computer Vision Applied Research Scientist at Boon, you will own end-to-end experiments on our foundation model, from architecture design through self-supervised pretraining, supervised fine-tuning, and shipping production models into our inference pipeline. This is a 50/50 research-to-production role. You will propose new architectures, run the experiments that prove or disprove them, and ship the winning models to real customers. You will have autonomy over direction and experimental ideas, staying aligned with the team and the company's research focus. This is not a role for someone who wants to be told what to build.

Requirements

3-7+ years in computer vision research (industry research lab, applied science team, PhD research + industry, or equivalent)
Strong track record of published CV research OR trained production CV models that shipped at scale
Hands-on expertise in multi-modal dense prediction (segmentation, detection, or joint vision-language tasks)
Production experience with modern vision transformer backbones (SAM, DINOv2/v3, CLIP, SigLIP, or similar)
Strong PyTorch fluency and experience training large vision models
Ability to move models from research to production inference pipelines
Strong fundamentals in deep learning: optimization, loss design, regularization, self-supervised learning
Fluency in English (written and verbal)

Nice To Haves

Experience with Graph Neural Networks or relational reasoning architectures (we do not expect this — most CV researchers do not — but it is a meaningful plus)
Experience with text spotting, OCR, or scene text detection integrated with vision models
Experience with LoRA, adapters, or parameter-efficient fine-tuning of large vision models
Experience with self-supervised pretraining (MAE, DINO, or similar)
Experience with engineering/technical drawings, document understanding, or layout analysis
Contributions to open-source CV research
Published papers at top venues (CVPR, ICCV, ECCV, NeurIPS, ICLR)

Responsibilities

Design and evaluate novel multi-stage vision architectures for construction drawing understanding — perception, text-object association, and relational reasoning across elements
Drive architecture decisions: backbones, decoders, fusion strategies, loss functions, training regimes
Run rigorous experiments with clean baselines, ablations, and held-out evaluation on real construction drawings
Own supervised training and self-supervised pretraining strategies
Pursue research directions that compound accuracy across trades and scopes
Take models from experimental notebooks to the production inference pipeline
Work hands-on with PyTorch, YOLO, SAM, DINO, and other modern CV stacks
Collaborate with ML engineers on deployment, quantization, and serving
Debug real failures on real customer drawings and close the loop into the next training run
Collaborate with the synthetic data, annotation, and infrastructure teams to make sure experiments have the data and compute they need
Partner with engineering leadership on the accuracy roadmap and strategic direction
Write clean internal research reports so the broader team can learn from your work
Present findings, trade-offs, and recommendations to engineering leadership
Help shape what data we acquire and annotate, based on what the model actually needs
Define evaluation datasets and metrics that track progress honestly — not Kaggle-style leaderboard chasing
Identify failure modes on real customer drawings and design experiments that address them