AI Engineer | Computer Vision + LLMs (NYC)

Deya; New York (NY); Hybrid; USD 150,000 - 230,000; Full time

Benefits offered by this job

Equity

Hybrid work arrangement

Job summary

Deya is hiring an AI Engineer to own the AI stack end-to-end in NYC. You will advance a computer vision pipeline for smartphone ocular imaging and build LLM systems to summarize patient history for clinicians, guiding check-ins and treatment adherence.

You should have production CV/ML experience, be fluent with modern LLMs, and own models in production. This hybrid NYC role includes equity and a strong data-driven clinical focus.

Qualifications

• Production experience across computer vision and machine learning with shipped systems.
• Hands-on with modern LLMs: prompting, evals, retrieval, tools usage.
• Strong engineering fundamentals; can own models in production.
• Ability to translate vague clinical questions into ML problems.
• Familiarity with AI tooling for development and evaluation.

Responsibilities

• Take ownership of the ocular image pipeline and extend it.
• Train, evaluate, and deploy vision models on proprietary clinical data.
• Define ground truth with optometrists/ophthalmologists.
• Address real-world capture issues: lighting, focus, skin tone, devices.
• Develop LLM systems for longitudinal patient data retrieval and summaries.
• Build robust evaluation infrastructure and regression testing.
• Implement guardrails and human-review paths for clinical claims.
• Manage model serving, latency, and on-device vs cloud tradeoffs.

Skills

Computer Vision

Machine Learning

LLMs

Prompting

Model Deployment

Evaluation

Ground Truth

On-device ML

Data Handling

System Ownership

Tools

Cloud Platforms

ML Frameworks

ON-device ML Tooling

Job description

AI Engineer | Computer Vision + LLMs (NYC)

NYC (Hybrid) · Remote Posted Aug 17, 2026

About the role

Own Deya's AI stack end-to-end: the computer vision work behind the largest smartphone ocular image dataset, and the LLM systems that reason over longitudinal patient care.

The dataset is the hard part. We already have it.

Most people working on AI in eye care are limited by data: small sets, clean lab captures, no outcomes attached, no way to see what happened to the patient next.

We have the largest dataset of eye images captured on a smartphone; it’s growing through active clinical pilots, and it’s tied to symptoms, treatments, and outcomes over time.

That’s the moat. This role is about turning it into clinical signal.

What we’re building

Deya is an AI platform that sits between doctor and patient, and captures symptoms, treatments, outcomes, and images over time to deliver remote patient care.

We already have the largest dataset of eye images captured on a smartphone and are currently piloting our platform with eye doctors and collecting real patient data.

The goal: build the largest longitudinal dataset in eye care (behavior + outcomes + ocular images) captured from connected devices, and use it to power AI systems that extend and improve clinical eyecare.

The role

You’d be the technical owner of AI at Deya, across two connected halves.

Vision. Our computer vision pipeline for smartphone-captured ocular imaging already exists and is in use. You’d take it over and take it considerably further; more robust models, trained and validated against our own clinical data, with inference that holds up on images captured by patients in their kitchens rather than by technicians in a clinic.

Language. The longitudinal record is only useful if something can reason over it. You’d build the LLM and agentic systems that summarize a patient’s history for a clinician in seconds, surface changes worth a doctor’s attention, and guide patients through check-ins and treatment adherence, with the evaluation discipline a clinical context demands.

This is a high ownership, high velocity role. You’d set the technical direction for AI here, working alongside the core engineering team and directly with our clinicians.

What you’ll work on

• Taking ownership of our existing ocular image pipeline and extending it
• Training, evaluating, and deploying vision models against our proprietary clinical data
• Ground truth definition, working directly with optometrists and ophthalmologists
• Handling the realities of consumer capture: variable lighting, focus, framing, skin tone, and device hardware
• LLM systems over longitudinal patient data retrieval, structured output, tool use, and agentic workflows for both patients and clinicians
• Serious evaluation infrastructure: eval sets, regression testing, failure analysis, and knowing when a model shouldn’t be trusted
• Guardrails, human review paths, and honest calibration of what the system does and doesn’t claim clinically
• Model serving, inference cost and latency, and on-device vs. cloud tradeoffs

How we build

AI is part of our daily development workflow, and for this role that goes further than tooling. Model capability is a moving target, and we want someone who treats keeping current as part of the job: reading the work, running new models against our own evals, and telling us when something we built six months ago should be rebuilt or thrown out. Bring your own view rather than taking published benchmarks at face value.

Who you are

• Production experience across computer vision and machine learning and shipped systems, not just notebooks and papers
• Hands-on with modern LLMs: prompting, evals, retrieval, tool use, fine-tuning where it’s warranted
• Strong engineering fundamentals; you can own your models in production rather than handing them off
• Comfortable working from an ambiguous clinical question to a defined ML problem
• Excited and comfortable using AI tooling for development
• Rigorous about evaluation and appropriately skeptical of your own results, which matters more when the output touches patient care
• Bonus: medical imaging, ophthalmic imaging, on-device ML, or clinical ML experience, EHR integrations
• Fast, scrappy, and high ownership

Why this is a unique moment

• A proprietary, growing clinical dataset most research groups simply can’t access, and you’d influence how it’s collected
• Vision + LLMs + longitudinal outcomes in a single system is genuinely underbuilt
• Clinicians in-house to define ground truth and validate what you build
• Because we control capture, you can improve the data itself, not just model around its limits
• Eye care is a massive, overlooked wedge in healthcare where AI hasn’t been meaningfully applied

This is not incremental. It’s a chance to help define the layer everything else will sit on.

Details

• Full-time
• NYC (hybrid)
• Competitive salary + equity