ML Research Scientist

About us

We're building an AI research-and-drafting assistant for Indian lawyers, on top of a legal
knowledge platform built from primary court sources. The hard part isn't generating text -
it's making sure every answer traces to a real source that is still good law. That's the problem
this role owns.

The role

You own how we find the right law and verify it. You'll build and compare the retrieval
approaches behind the product - how documents are embedded and searched, how a
citation graph is modelled and traversed, and how we check that what the system says is
backed by a real source. Every choice is settled on evidence against a legal eval set, not
intuition, so you'll work closely with our in-house lawyer for ground truth.
This is a hands-on research role on a small team: you run the experiments and make the call.

What we're looking for

• Strong hands-on experience with embeddings and modern retrieval - dense and
• lexical search, rerankers, and the tradeoffs between them.
• Work with graph databases and knowledge graphs: modelling entities and
• relationships, and using graph traversal as part of search.
• Eval-driven - you reach for an experiment before an opinion, and you can design a
• benchmark that measures what matters.
• Comfort where learned models meet symbolic or rule-based checks.
• You write clearly enough that an engineer can build from your spec.

What it takes to join us

• We are an early-stage startup, and that shapes who does well here:
• You own problems end to end - from the open question to the experiment to the
• call the rest of the team builds on. Nobody hands you a spec.
• You're comfortable without a playbook. Much of this is unsettled, and you'd rather
• scope the ambiguity than wait for someone else to.
• You move fast and decide on the evidence you have, then revisit when better
• evidence shows up
• You want the range and the say that come with being early, and you're fine with the
• rough edges that come with it.

Nice to have

• Legal, regulatory, or another high-stakes domain where being wrong has
• consequences.
• Familiarity with Indian law or Indian-language NLP.
• Extraction / NER experience (citations, entities from documents)