Everyone wants to work on AI in healthcare. Almost nobody gets to build it. Recollia builds AI for medical students: tutors, clinical cases, exam questions, image Q&A;, and the medical images behind all of it, generated with AI, checked by doctors, used by students every day. You'll own the generative side of that.

WHERE THIS GOES

This starts as an internship with a stipend of Rs 20,000 and a possible incentive of Rs 5,000 every two weeks, depending on performance. Make our generated images, cases, and questions good enough that doctors sign off on them and students learn from them, and it converts into a founding engineer role with equity. The bar is high on purpose: a wrong image in medicine teaches the wrong medicine.

WHAT YOU'LL BUILD

- Medical image generation: anatomy, histology, pathology, radiology-style teaching images, mechanism diagrams; fine-tuned diffusion models that get the anatomy right, at a cost per image we can afford at scale
- Clinical cases and MCQs: exam-grade vignettes and questions (NEET-PG, USMLE and similar), grounded in sources, structured, and reviewed before a student sees them
- Image Q&A;: a multimodal tutor that can look at an X-ray, ECG, or slide and teach from it
- The corpus: a versioned, licensed, de-identified medical education dataset of images, text, and questions; the asset Recollia compounds on
- Evals and review loops that prove the medicine is right, with doctors in the loop and their corrections feeding back into the models

REQUIRED

Generative images

- Diffusion models beyond the prompt box: fine-tuning (LoRA, DreamBooth or similar), ControlNet, inpainting, consistent style across thousands of images
- Anatomical accuracy as a hard constraint:



you know why a six-fingered hand is funny and a mislabeled artery is not
- Evaluating generated images past FID: expert review, failure taxonomies, and a rejection rate you can quote
- Fairness built in: dermatology on every skin tone, anatomy across ages and sexes

Generative text

- Clinical case and question generation with LLMs: structured outputs, source grounding with citations, hallucination control
- Item-writing craft: what makes an MCQ fair, what makes a distractor work, how to tell a recall question from a reasoning one (NBME-style guidelines, Bloom's levels)
- Measuring question quality with real students: difficulty, discrimination, and retiring the items that fail

Multimodal

- Vision-language models (Qwen-VL, LLaVA, MedGemma or similar): fine-tuning on image-text pairs, evaluation on medical image Q&A;, knowing where they confidently fail
- Building image-text datasets: captioning, labeling, review, and the licensing to use them

Data and corpus

- Dataset pipelines: ingestion, cleaning, dedup, de-identification, versioning; data you can trust and reproduce
- Synthetic data with expert review, and knowing when synthetic is the wrong answer
- Licensing and consent for medical content; no patient data without the right to use it

Evaluation and safety

- Medical accuracy evals with clinicians, LLM-as-judge and its failure modes, regression suites




- Every generated image labeled as synthetic and educational; nothing generated is ever passed off as a real patient
- Cost per image, per case, per question, against the API baseline

Medicine

- No medical background required, but real curiosity about how medicine is learned: how a second-year gets through anatomy, why a radiograph is hard to read, what a valuable examiner's question does
- You'll sit with doctors and students, learn enough medicine to know when the model is wrong, and build the loop that lets them tell you

Engineering foundations

- Python and PyTorch at production level: diffusers, transformers, async pipelines, tests
- GPU work: fine-tuning on rented GPUs (RunPod, Modal, AWS), quantization, memory math, cost per asset
- FastAPI or equivalent, Postgres, S3, Docker, Git, CI; experiment logs (W&B; or equivalent)

No degree required. Skills over credentials.

WHAT HUNGER LOOKS LIKE TO US

- You've generated a thousand images and kept a hundred, and you can say why.
- You'd rather show a doctor the output than debate it.
- A wrong answer in a medical MCQ bothers you until it's fixed.
- You ask for more scope, not more instructions.
- You say when something is broken, including our decisions.

HOW TO APPLY

Email [email protected] with the subject "Generative AI, Healthcare" and one thing you built with generative AI, whether image, text, or multimodal: repo, demo, or write-up. In three sentences, tell us what it does, what broke, and what you'd do differently. Then one more thing: generate one medical teaching image with any tool you like, attach it, and tell us what a doctor would correct.

📌 Generative AI Engineer, Healthcare (Founding Track) (India)
🏢 Recollia
📍 India