Research Engineer, Post-Training

Research Engineer, Post-Training

Location:

Bengaluru, India

Type:

Full Time

Team:

Engineering / AI Research

Why Jurisphere

Jurisphere is building AI systems that can do serious legal work.

We recently raised a $2.2M seed round led by Info Edge Ventures

, with participation from Flourish Ventures, Antler, and 8i Ventures. Today, Jurisphere is used by 500+ organisations and 20,000+ legal professionals

, including ICICI Bank, Unilever, Philips, Tata Capital, and CMS IndusLaw.

We are a small team working at the intersection of frontier AI, legal reasoning, and high-stakes professional workflows

.

The Role

We’re hiring a Research Engineer, Post-Training

to help make our models meaningfully better at legal work.

You will work across model training, datasets, evaluations, agents, and expert feedback — taking problems from an observed model failure through to an experiment, training signal, and measurable

improvement.

This is a hands-on research engineering role for someone comfortable working with open-weight models and ambiguous research problems.

What You’ll Do

• Run post-training experiments across SFT, preference optimization, RLHF/RLAIF, reward modelling, distillation, and related techniques

  .

• Build training datasets from expert feedback, model traces, legal workflows, and synthetic data

  .

• Design evaluations and reward systems for legal reasoning, research, drafting, document analysis, and citation accuracy.

• Study model and agent behaviour to identify systematic failure modes.

• Build agent environments involving retrieval, tools, subagents, validation loops, and long-horizon workflows

  .

• Turn research findings into better training data, model recipes, evaluations, and production systems

  .

• Work closely with engineers and legal experts to translate expert judgment into scalable systems.

What You Have

• Hands-on experience training or post-training LLMs

  , particularly open-weight models.

• Strong Python and ML engineering

  skills.

• Experience with some combination of SFT, preference learning, RL, reward modelling, distillation, agents, or evaluations

  .

• Strong experimental judgment: you can formulate hypotheses, run controlled experiments, analyse model behaviour, and interpret ambiguous results.

• Ability to independently own technically difficult, loosely defined problems.

Nice to have:

distributed training or GPU experience, agent systems, evaluation infrastructure, research publications, open-source contributions, or experience in high-stakes domains such as law, finance, or healthcare.

Why This Role

Post-training at Jurisphere is the feedback loop between models, real legal work, expert judgment, and engineering

.

You’ll work across the full loop: understand why a model failed, build an evaluation that captures it, create a training signal to improve it, and measure whether the next model is actually better.

The goal: build models that are meaningfully better at doing legal work.