ML Engineer - Runtime

About the Company Adalat AI is building an end-to-end justice tech stack that automates manual and clerical pain points in courtrooms, giving judges back time to focus on what matters most: decision-making and delivering justice. Our solutions - from AI-powered transcription in Indian languages to case-flow management and document navigation - are now deployed across 9 states, covering nearly 20% of India’s judiciary. Backed by leading technology companies and funders, and incubated at MIT and Oxford, Adalat AI is working to eliminate judicial delays and expand access to timely justice. Founded by a team with backgrounds in law, technology, and economics from Harvard, Oxford, MIT, and IIIT Hyderabad, we are scaling rapidly across India and the Global South. Role Overview Most ML systems are built for controlled environments. Adalat's are not. Our models run across 5,000+ courtrooms — from the Supreme Court to remote district courts with unreliable connectivity — processing 8–12 hours of live audio every day. A model that works in a data centre has to survive rooms with ceiling fans, simultaneous speakers in three languages, and hardware that hasn't been refreshed in a decade. That gap between the lab and the courtroom is what this role owns. In this role, you'll be responsible for making our speech and language models fast, cost-effective, and deployable at scale on real-world hardware. You'll work at the intersection of model architecture and systems engineering — applying compression techniques, building compiler-aware inference pipelines, and designing serving infrastructure that holds up under the most demanding conditions in Indian legal AI. This is not model-agnostic infrastructure work: you'll collaborate closely with our Speech and NLP researchers to understand what each model does and find the optimisations that preserve what matters while cutting what doesn't. Key Responsibilities

1. Optimise inference pipelines

• Apply quantisation (PTQ, QAT), pruning, and distillation to reduce model size and latency without meaningful quality loss.
• Build compiler-aware workflows using MLIR, LLVM, or ONNX to squeeze performance from diverse CPU and GPU targets.
• Profile, benchmark, and continuously improve throughput across the full ML stack.

1. Build and maintain scalable serving infrastructure

• Design async batching, concurrency, and load-balancing strategies for high-throughput, real-time audio processing.
• Architect deployment pipelines that span cloud and edge environments — including constrained, low-bandwidth courtroom settings.
• Own cost efficiency: understanding exactly where compute is being spent and driving it down without degrading output quality.

1. Debug and harden production systems

• Stress-test inference across thousands of hours of real courtroom audio and documents.
• Build diagnostic tooling for runtime failures — latency spikes, memory pressure, model drift in noisy acoustic environments.
• Create monitoring and alerting infrastructure that catches regressions before users do.

1. Partner with Research and Product

• Translate system constraints back into model architecture recommendations for Speech and NLP researchers.
• Work with the Product ML team to define latency and cost targets that are both technically achievable and user-meaningful.
• Contribute to model design decisions early — before a model is trained in a way that makes efficient deployment impossible.

Qualifications Must have

• 3–6 years in ML optimisation, inference engineering, model compression, or compiler development.
• Strong Python and C/C++ programming.
• Hands-on experience with quantisation, distillation, or pruning in production.
• Proficiency with deep learning frameworks (PyTorch or TensorFlow) and at least one compiler framework (LLVM, MLIR, ONNX, TensorRT).
• Experience deploying ML models in resource-constrained or latency-sensitive environments.

Strong plus

• Experience with advanced LLM serving strategies (vLLM, continuous batching, speculative decoding).
• Prior work with speech pipelines (ASR, diarization) at the systems level.
• Experience on edge or embedded hardware targets.
• Publications at systems or ML venues (MLSys, OSDI, NeurIPS, ICML, Interspeech, or similar).

What You Will Achieve in a Year You'll have built models for legal QnA, search, summarisation, and translation that work on real Indian legal documents — and you'll have the eval methodology to prove it. You'll have made meaningful progress on document intelligence: OCR and structured extraction from the kinds of scanned, degraded documents that existing tools give up on. You'll have at least one paper submission in progress. And the benchmarks you defined will be the ones the team uses to evaluate everything that comes after. Join Our Team If you're driven by meaningful work and want to build technology that changes how justice is delivered — we'd love to hear from you. Apply now and be part of a team that ships fast, thinks deeply, and works on problems that matter. 👉 Apply here