ML Research Scientist Mumbai · 5+ years · Full-time →

NyaayOS; Mumbai; On-site; INR 3,000,000 - 6,000,000; Full time

Job summary

ParamAIQ is seeking a research-focused engineer to own end-to-end representation, retrieval, and grounding of rapidly evolving legal knowledge. You will design data representations, build retrieval pipelines, and verify outputs against real sources to ensure authority.

You will lead experiments, create benchmarks, and integrate LLM components with symbolic logic to produce reliable, grounded answers in a high-stakes information domain. Strong Python and experiment design are essential.

Qualifications

• Experience building and evaluating retrieval systems with an emphasis on authority.
• Ability to design benchmarks and measure factual grounding in legal knowledge.
• Proficiency in Python and building end-to-end ML/RL pipelines.

Responsibilities

• Own how we represent a large, changing body of law for machine reasoning.
• Retrieve the right law with precision where a single fact matters.
• Verify that answers are grounded in real, valid sources and maintain freshness.
• Design experiments and benchmarks; own the evaluation metrics.

Skills

Retrieval systems

Knowledge representation

Experiment design

Python programming

LLM integration

Grounded reasoning

Tools

Dense embeddings

Vector search

RAG pipelines

Job description

Generating text is the easy part. The hard part is making sure every sentence we put in front of a lawyer rests on real, currently valid authority. That is the problem you own. You own how ParamAIQ represents legal knowledge, and how it finds and verifies the law behind an answer. Indian law is vast, cross-referenced, and always moving statutes get amended, judgments get overruled, positions shift. Turning that into something a machine can search, reason over, and stand behind is the core research problem of the company. You settle it with experiments against a legal eval set you build with our in-house lawyer, not with intuition. The work ships into a product lawyers rely on, and it is judged there.

YOU WILL OWN

• How we represent a large, changing body of law so a machine can reason over it and know what is currently authoritative.
• How we retrieve the right law with precision, where the answer can turn on a single fact, provision, or jurisdiction.
• How we verify that an answer is grounded in a real source that is still good law and catch it when it isn't.
• The legal eval set as the source of truth. You design the benchmarks and you own the number.
• The research direction. On a small team, you frame the problem, run the experiment, and make the call.

THE HARD PROBLEMS

• Representing law that changes under you. An authority that governs today can be overruled tomorrow, and the representation must know the difference.
• Proving groundedness at scale. A fluent answer citing a real sounding but wrong or dead authority is worse than no answer. Verifying grounding reliably is unsolved in this domain.
• Defining "correct" where experts disagree. Before you can measure a system, you must build ground truth in a field where correctness is argued, not given.

WE ARE LOOKING FOR

• Strong command of modern retrieval and representation: dense and sparse embeddings, vector and lexical search, rerankers, and when each win. You can build and evaluate a retrieval system end to end.
• Hands-on at the model layer for RAG and reasoning: prompting, structured extraction, and using LLMs as components of a larger system.
• Knowledge representation in practice: knowledge graphs, entity and relationship modelling, and using structure inside search and reasoning.
• Eval-driven for real: you design benchmarks, build datasets, and measure factuality and groundedness, not just offline accuracy.
• Comfortable where learned models meet symbolic or rule-based logic.
• Strong Python and rigorous experiment design, and you write clearly enough that an engineer can build from your spec.

NICE TO HAVE

• A high-stakes domain where being wrong has real consequences: legal, medical, financial.
• Indian law, or Indian-language NLP.
• Information extraction and NER from messy documents.

THIS ISN'T THE ROLE IF

• You want a narrow lane. You own the problem end to end: how law is represented, how it's retrieved and verified, and the eval that decides what ships.
• You want clean, settled ground truth. Defining "correct" in a contested field is part of the research, not a prerequisite for it.