Co-founder and head of machine learning

The role

Zohardi is looking for a co-founder to lead machine learning. You'll own the models that turn every flight into findings about the land, including how they learn each new property, how they're evaluated and how they improve safely as we grow.

You'll lead the modeling and training loop for a computer vision system that learns from aerial and remote sensing imagery over time. It has to learn each new property quickly, keep what it already knows and spend people's time only where it adds the most. This is a founding role. You'll set the technical direction for machine learning, build the team around it and help shape the company with the founding team.

About Zohardi

Zohardi is a geospatial intelligence company in Santa Barbara, California. We started with the question every landowner and land manager asks. What changed out there, and how sure are we?

A Zohardi Brain, our on-site dock, flies aircraft on a schedule and keeps a living digital twin of each property up to date. We own and operate every dock and aircraft, and we obtain and maintain the FAA approvals to fly beyond visual line of sight. Farms, ranches, preserves, corridors and waterways ask different questions of the same ground, and one set of flights serves them all. Results land in Esri ArcGIS and Databricks.

We're a small team of environmental scientists, statisticians, data scientists and computer vision engineers, building docks, aircraft, models and a data platform built for real sites. We're an Esri partner, a Databricks partner and an NVIDIA Inception member.

The data you'll work with

Your models will learn from the same ground, flown again and again from the same dock, at about 2 cm per pixel. Each flight brings visual, multispectral, thermal and gas data, read together with the dock's own weather record. Every flight is processed on the dock, so findings arrive the same day, and every finding carries its confidence and the evidence behind it.

What you'll own

• Training data. Build the labeling and review loop behind our training data, find the gaps in its coverage and close them.
• Human review. Decide which work needs a person, keep label quality measured and focus human effort where it adds the most.
• Compact models. Train compact, real-time computer vision models that run on the dock, and measure how close they come to the best results available.
• Adaptation and continual learning. Adapt models to each new property and season without starting over, prevent forgetting and gate every release with rollback.
• Change detection. Flag what changed on a property between flights, with calibrated false-alarm rates and drift monitoring.
• Evaluation. Locked held-out test sets, success criteria set before experiments, results across several seeds, and findings that drive decisions.
• Deployment. Register, version and serve models with MLflow, optimize them for the dock and the cloud, and release them through automated quality gates with rollback.
• Leadership. Make and document modeling decisions, direct the engineering team and review their evidence. As a co-founder, you'll also hire the machine learning team and shape the product and roadmap with the founding team.

The models you'll work with

The field moves quickly and our choices will change, so experience with comparable models counts the same. What matters is fluency with these model families and the methods behind them.

• Vision backbones and self-supervised embeddings
• Foundation models for detection and segmentation
• Real-time detectors, DETR-family and YOLO-class
• Parameter-efficient fine-tuning and adapter methods
• Anomaly and change detection
• Vision-language models
• Taxonomic models for species identification
• Edge inference with TensorRT or ONNX Runtime on embedded hardware

What success looks like in your first year

• A repeatable labeling and training loop that measurably reduces the human effort behind each model
• Production models chosen on evidence, with commercial licensing cleared
• Model updates released through automated quality gates, without regressions
• Change detection running on real data, with a stated false-alarm rate
• A measured cost, in data, compute and time, to add a new kind of finding or a new property

Must have

• 7+ years of applied machine learning, including 3+ years owning computer vision models in production
• Has shipped a system that trains on model-generated labels, such as pseudo-labeling, knowledge distillation or auto-labeling pipelines
• Hands-on with foundation vision models for detection or segmentation
• Has trained and shipped modern real-time detectors, DETR-family or YOLO-class
• Parameter-efficient fine-tuning, such as LoRA or adapters, and continual learning, including catastrophic forgetting, replay and regression testing
• Has built human-in-the-loop labeling with active learning or uncertainty-based routing, and quality audits
• Self-supervised embeddings used for retrieval, anomaly or change detection, including memory-bank methods
• Strong statistics, including calibration, conformal prediction, sampling design, and variance and significance across runs
• Strong PyTorch, and able to turn a new paper into a measured baseline quickly
• Confident daily use of AI coding harnesses such as Claude Code, Codex or Cursor, and multi-agent orchestration where one agent builds and another reviews, to ship quickly without losing precision
• Experiment tracking and a model registry with MLflow or equivalent, with versioned models, clear lineage, staged promotion and rollback
• Diligence on model licenses and data-use terms

Nice to have

• Drone, aerial or satellite imagery, including orthomosaics, georeferencing and image registration
• Wildlife, agriculture, ecology or land management applications
• Taxonomic models for species identification
• Multispectral or thermal data
• Edge optimization with TensorRT or ONNX Runtime on embedded hardware
• Databricks, including Unity Catalog, Delta Lake, Lakeflow Jobs, Model Serving and GPU training on the platform, with Azure ML also useful
• Publications, open-source work or a documented negative result

How you work

• Set success criteria before running experiments
• Write decisions down and revisit them when evidence changes
• Prefer an honest "inconclusive" to an overstated win
• Review evidence, not summaries, whether it comes from a teammate or an agent

What we offer

• The co-founder title, with salary and equity
• The chance to build machine learning at Zohardi from the start, and the team around it
• Work measured in the field, on working land
• Remote within the United States

How to apply

Apply below with your resume and a cover letter, each a PDF of 2 MB or less. After a first conversation, we'll share the full technical profile for the role.

The role is fully remote, with occasional travel to Santa Barbara and to the properties we fly. You must be authorized to work in the United States. We can't sponsor visas for this role.

Zohardi is an equal opportunity employer. We consider all qualified applicants without regard to race, color, religion, sex, sexual orientation, gender identity, national origin, age, disability, veteran status or any other characteristic protected by law.