ML Research Engineer - Post-training (LLMs)

ML Research Engineer; Post-training (LLMs)

Bengaluru · Full-time · Experience: 4-8 yrs

India's healthcare runs in twenty-two languages, on handwritten prescriptions and ten-minute consults, and the models that should serve it are trained on the English internet. We're fixing that, in the open.

About EkaCare

EkaCare is India's connected healthcare platform: an EMR that doctors run their practices on, a personal health record used by millions of Indians, and one of the deepest integrations with India's ABDM digital-health rails. Our Parrotlet family of medical models already serves Indian doctors in production, and we open-source our work where it counts.

The role

Post-training is where a base model becomes a doctor's tool, and where most medical models quietly fail. You'll turn a strong pre-trained base into a model that follows instructions, knows what it doesn't know, stays safe in a clinical setting, and does it in a dozen Indian languages.

What you'll do

• Design SFT/IFT data mixes and chat templates for clinical tasks, and reformat the world's medical data to match.
• Run preference optimisation (DPO/ORPO/GRPO-class) and reward-model training; own the ablation grid.
• Build RL loops with verifiable medical rewards, with practising physicians in the loop. Real doctor-in-the-loop, not proxy labels.
• Create agentic and tool-use training data for healthcare workflows.
• Live in the eval → error-analysis → iterate loop; red-team your own model before the world does.

What we look for

• 2-4 years in ML with hands-on post-training of ≥7B open-weights models; you've shipped SFT plus at least one preference-optimisation method end to end, not a notebook demo.
• Fluency with the open post-training stack (TRL / NeMo RL / OpenRLHF; vLLM for rollouts).
• Strong empirical taste: ablation discipline, LLM-judge literacy, contamination paranoia.
• Solid engineering, Python, distributed-training basics, comfort in a fast codebase.

Bonus

• PPO/GRPO at scale; reward-hacking war stories.
• Multilingual or medical/clinical alignment work.
• Open-source contributions people actually use.