ML Research Engineer - Legal Reasoning

Role Overview The Justice Lab & Tax-Litigation Co-Pilot Youll join The Justice Lab —our skunk-works unit focused on ideas six-to-twelve months ahead of production. Flagship project: a Tax-Litigation Co-Pilot that ingests full judgments, statutes, and filings, then produces defensible predictions and transparent explanations. Project Coordinator: Arghya Bhattacharya (CTO, Adalat AI) Project Oversight: Prof. Daron Acemoglu (Nobel Laureate, MIT) & Prof. Daniel Kang (UIUC) Role in a Nutshell As a Research Engineer—Legal Reasoning you will turn cutting-edge ideas into artifacts that ship: Frame the problem: formalize legal reasoning for outcome prediction. Design experiments: benchmark LLMs on labeled tax-law and civil-procedure tasks. Prototype systems: retrieval-augmented generation, evidence tracing, causal inference—pipelines that think like lawyers . Build eval suites: factual consistency, citation faithfulness, policy impact (e.g., case-load reduction). Ship hand-offables: lightweight services or notebooks that engineers can harden. Publish: co-author internal memos and external papers with academic partners. Key Responsibilities Data & Evaluation Curate, label, and version corpora spanning four court tiers. Create task sets for prediction, entailment, and explanation. Modeling & Experimentation Fine-tune / distill LLMs with RL-, DPO-, or SFT-style feedback. Explore long-context and retrieval strategies (LoRA, RAG, chunking). Legal-Reasoning Research Model precedential hierarchies, detect conflicts, and generate citation-grounded chains of thought. Collaboration Sync daily on design and code quality. Present findings to Professors Acemoglu, Kang, and policy advisors. Documentation & Dissemination Maintain reproducible logs, polished reports, and publish-ready code. Qualifications Must-Have Nice-to-Have 2 + years NLP/ML research (industry or grad school) Prior work on legal or policy datasets Fluency in PyTorch/JAX & modern LLM fine-tuning stacks Publications at ACL, ICML, NeurIPS, etc. Skill in large-corpus wrangling & eval pipeline building Causal-inference or decision-theoretic ML Clear, concise technical writing & comms Familiarity with Indian tax or civil-procedure law No one ticks every box—if the mission resonates, lets talk.  What You Will Achieve in a Year A prototype that classifies appeal merit with 75 % F1 on held-out High-Court cases. An evaluation methodology poised to become the standard for legal-AI outcome prediction in the Global South . A first-author or co-author paper submission (e.g., NeurIPS L4DC, ICML LawML). Pilot deployment inside real-world Tax Offices.