Remote-first (India)
India has more than 50 million pending court cases. Behind that number are people waiting years for decisions that affect their lives. Part of the problem is that courts still spend enormous time on manual and clerical work: recording proceedings, preparing orders, moving documents, and keeping fragmented systems in sync. Judges should be spending that time delivering justice.
Our mission at Adalat AI is to make timely justice possible. We are a Y Combinator–backed nonprofit building an end-to-end technology stack for justice systems. We want to do for courts what UPI did for payments: create a common, reliable technology layer on which an entire ecosystem can work. That means going beyond individual AI features to build the infrastructure through which courts can operate more efficiently, securely, and accessibly.
Our products already span AI transcription in Indian languages, dashboards for judges and court administrators, WhatsApp assistants for litigants and advocates, AI-powered workflows for courts and registries, and a Paperless Courts platform for state judiciaries. They are used in more than 5,300 courtrooms across 10 Indian states—roughly 25% of India’s judiciary. Our next ambition is to reach every courtroom in the country, and eventually help justice systems across the Global South.
Our work has been featured in a Stanford case study and The Wall Street Journal, and recognised by the Harvard President’s Innovation Challenge, MIT Solve, and the AI Innovation Challenge at the Global AI Summit. Incubated at MIT and Oxford, we are supported by leading technology companies and funders.
Founded by a team with backgrounds in law, technology, and economics from Harvard, Oxford, MIT, and IIIT Hyderabad, we are now building the teams and technology needed for that next stage of scale.
Role Overview
Why this role matters
The company-wide problem is making justice delivery faster by removing manual work across the lifecycle of a case. Transcription and dictation were our initial wedge: practical tools that reduced clerical work, helped courts build trust in our technology, and opened the way to an end-to-end technology stack for courts.
Speech remains a foundational part of that stack. A judge dictating an order, a lawyer making a submission, or witnesses giving evidence all need an accurate, usable record. Reliable transcription and diarization reduce the time spent creating and correcting those records and make them available to the workflows that follow.
But Indian courtrooms are among the hardest environments for speech technology. A judge may dictate in Malayalam, switch to English for a statutory reference, and name a village the model has never heard. Lawyers interrupt one another; people speak from across a noisy room; proceedings can involve many speakers. Models built for clean, consumer speech do not reliably handle this diversity. We need ASR and diarization that work across Indian languages, accents, code-switching, legal vocabulary, microphones, and real courtroom conditions—not just on benchmarks.
Today our speech products already reach courts at significant scale. Getting to every courtroom requires another step change: stronger models, more languages, better evaluation, lower inference costs, dependable deployments, and a larger team that can ship improvements quickly.
We are hiring a Head of Speech Technology as a foundational leader in this next phase of team-building. You will lead dictation and diarization research and engineering, establish the technical direction, and help turn advanced speech ML into reliable infrastructure for courts across India.
The role
You will build and lead the speech ML team across our dictation and diarization pods. You will directly manage speech researchers and speech ML engineers—not the runtime engineers or general software engineers who work alongside the pod. You will own the speech research agenda, model quality, the team’s growth and standards, and the path from promising experiments to products deployed in courts. You will work closely with runtime, platform, and product engineering leaders to make those models fast, affordable, and reliable in production.
We are looking for someone who has helped build speech technology at serious scale—at organisations behind products such as Alexa, or comparable speech platforms—and who understands both frontier research and what it takes to ship a product people use every day. You should be as comfortable discussing a model architecture or a top-tier paper as you are making a difficult product trade-off, coaching researchers, or diagnosing an issue from the field.
This is a fast-moving, high-ownership environment. Requirements evolve as we enter new courts and languages; teams cannot wait for perfect data or complete specifications. You will need to set a high technical bar, make decisions with incomplete information, and help a growing team move quickly without compromising accuracy, reliability, or user trust.
You will work closely with our founders, research and engineering leadership, product designers, language and data teams, and colleagues who work directly with courts. This is an opportunity to help shape both the speech technology and the research organisation we need to serve the whole country.
What success looks like
Within your first six months, you will have built a clear roadmap for multilingual courtroom ASR and diarization, set up an effective speech research and engineering team with explicit ownership, and shipped measurable improvements in priority languages and courtroom workflows. You will have established evaluation and release practices that make model quality visible beyond aggregate benchmarks, and a practical approach to adding new languages and deployments without allowing cost or operational complexity to outgrow adoption.
Over time, success means a speech ML organisation that can continuously adapt to India’s linguistic diversity and courtroom conditions, and a research-to-product system that scales beyond the founding team. We will judge progress by the quality of the team you build and what courts experience: fewer transcript corrections, dependable speaker attribution, reliable service during court hours, and speech technology affordable enough to reach every courtroom.
Key Responsibilities
Speech strategy and execution. Set a multi-year research and product roadmap for courtroom ASR, dictation, and diarization across Indian languages. Translate field failures into research problems, model milestones, and release decisions. Make deliberate trade-offs between accuracy, latency, reliability, compute cost, and speed of delivery as we expand toward every court.
An exceptional speech ML team. Recruit, mentor, and develop speech researchers and speech ML engineers; build strong technical leads; set clear ownership; and create a culture of rigorous experimentation and fast execution. This is an early leadership hire in a broader expansion of Adalat AI’s ML and engineering organisation.
ASR for every courtroom. Build and adapt streaming and batch ASR across Indian English and Indic languages, including code-switching, regional pronunciation, legal vocabulary, unfamiliar names, and document-ready punctuation and numerals. Develop methods that generalise to new languages and low-resource conditions, and improve robustness to noise, reverberation, distant microphones, and uneven recording quality.
Diarization for real hearings. Lead systems that attribute speech consistently through long proceedings, overlapping turns, interruptions, and changing speaker counts. Bring diarization and ASR together into transcripts people can review and use.
Data and evaluation that guide decisions. Work with language and data teams on collection, annotation quality, adaptation, and efficient learning in low-resource settings. Build representative evaluations across accents, languages, court settings, and hardware. Track consequential errors and correction effort alongside WER, diarization metrics, and speaker-attributed transcription quality.
Research that survives production. Partner with runtime, software, and platform teams on the path from model to service: streaming inference, quantisation and optimisation, concurrency, observability, versioned releases, and rollback. You will own speech-model decisions and product quality, while collaborating rather than directly managing the runtime and software engineering teams. Help make cloud and on-premises deployments work within courtroom compute, connectivity, privacy, and data-residency constraints.
A disciplined research-to-product loop. Stay close to users and failure cases. Decide when to adapt an existing model, invest in better data, improve the surrounding system, or pursue a new research direction. Turn useful research into shipped capability and publish or open-source work where appropriate.
Qualifications
At least 10 years of relevant experience, with substantial depth in speech ML and a record of taking speech systems into production.
At least 3 years of direct people management, including leading speech ML projects at a speech-focused company or an established speech organisation. You have hired, coached, and evaluated researchers or engineers and owned delivery through multiple releases.
Deep technical judgment in speech recognition, with meaningful experience in diarization, speaker modelling, or closely related speech and audio problems. You can explain the modelling, data, and systems decisions behind the products you shipped.
Production ownership. Evidence that your team improved a deployed speech system under real latency, reliability, or cost constraints. You understand how streaming inference, deployment architecture, and model quality affect one another.
Hands-on fluency. Strong grounding in modern speech architectures, training and evaluation, Python, and a framework such as PyTorch. You remain able to inspect experiments and engage credibly in code and design reviews.
Research credibility. A strong record of original technical work, ideally including publications at top speech or ML venues such as Interspeech, ICASSP, NeurIPS, ICML, or ICLR. You know how to choose hard research problems and turn promising results into better products.
Education. A B.Tech or equivalent undergraduate degree and a master’s degree in computer science, electrical engineering, machine learning, or a related discipline. A PhD is strongly preferred.
Leadership across disciplines. You can lead speech researchers and ML engineers, collaborate effectively with separate runtime and software engineering teams, make difficult prioritisation decisions, and build a team that operates with increasing independence.
Comfort with pace and ambiguity. You thrive in an ambitious, fast-moving environment where data is imperfect, production requirements change, and the team has to learn directly from users.
A reason to care about the problem. You want to apply your experience to public infrastructure and are willing to learn from the people who use it every day.
Experience with Indic or other low-resource languages, multilingual ASR, noisy or far-field audio, model compression, GPU inference, or on-premises deployments would be especially valuable. Substantial contributions to widely used speech products or open-source speech systems are especially compelling.
Know more about Adalat AI
Why Adalat AI
Some problems are technically hard. Others have the potential to change millions of lives. It is rare to work on one that is both—and rarer still to do it alongside people who care as deeply about the mission as you do.
At Adalat AI, we are trying to build the infrastructure that makes justice work at scale. We believe courts can have their UPI moment: a shared technology layer that transforms how an entire system operates. If we succeed, the work we do today in Indian courtrooms could become part of public infrastructure serving billions of people over time.
Getting there demands serious research and engineering. Courtroom speech combines some of the hardest problems in ASR and diarization: multilingual and code-mixed speech, low-resource languages, many speakers, noisy environments, and real deployment constraints. Our team has already released Koyal, open models for Indic dictation, and SCRIBE, an evaluation toolkit built to reveal errors aggregate scores can hide. There is much more to build.
You will work with researchers, engineers, designers, lawyers, and people close to the courts who share an ambition to make justice faster and more accessible. This is a place to do difficult work, move quickly, learn from real users, and see the results matter outside a benchmark or a product dashboard.
If you have ever wanted to use your technical skills at the intersection of hard problems, exceptional teammates, and meaningful impact, Adalat AI is a place to do it.