The Problem and Context
India faces a severe justice crisis with 50 million pending cases that would take 300 years to clear at current rates. Three out of four prisoners are simply awaiting their hearing dates, languishing in jail while their families wait in courts. Indian courtrooms are overwhelmed with paper—towers of rotting documents, typewriters instead of computers, and judges handwriting every word by hand because 90% of courts lack stenographers. Justice has become a question of logistics rather than law, with “the process itself has become the punishment.”

Adalat AI’s Solutions
Adalat AI is a legal tech nonprofit building AI solutions to address four critical courtroom bottlenecks. They’ve developed AI-powered transcription that understands legal jargon in multiple Indian languages and dialects, a Salesforce-like case management system to help judges efficiently navigate 50-100 daily cases, LLM-driven document digitization and analysis tools, and WhatsApp-based case updates for citizens. All solutions are built in-house with state-of-the-art encryption and deployed on Indian servers to meet strict security requirements. The team includes lawyers who provide extensive training programs covering not just how to use Adalat AI, but fundamental digital and AI literacy.

Impact and Future
Operating in 4,000+ courtrooms across 9 Indian states (10-15% of the judiciary), Adalat AI has demonstrated 2-3x increases in judicial productivity through randomized controlled trials, potentially reducing case resolution timelines by 30-50%. Kerala state has made Adalat AI mandatory statewide starting November 1st—the first judiciary-mandated AI system globally. The nonprofit is expanding to African countries (Ghana, Nigeria, Zambia) that share similar colonial legal structures and sustaining operations through three revenue streams: selling to private law firms and arbitration centers, government funding as budgets adapt, and international sales to wealthier countries while cross-subsidizing impact in the global south.

Adalat AI
https://www.adalat.ai/

Arghya Bhattacharya
Co-founder and Chief Technology Officer
Arghya Bhattacharya, co-founder & CTO, is a leading AI and machine learning engineer. He holds a bachelor’s in computer science and a master’s in artificial intelligence from IIIT Hyderabad, with publications in leading journals like ACL, CoNLL, and EaMT. As the first founding engineer at Enterpret, he led the development of cutting-edge NLP models, translating AI innovations into practical solutions. His expertise spans AI-powered identity verification, ML systems for automation, and large-scale AI product development.

Full Transcript of Presentation

Roland Vogl: Arghya Bhattacharya is the co-founder and CTO of Adalat AI . He’s working on an amazing project which will revolutionize, first of all, the legal system in India, but then eventually globally. We really appreciate you joining us here today, for our CodeX group meeting. 

Arghya: Thank you so much, Roland, and hello, everyone. It’s a pleasure being here. 

We’re building Adalat AI, but before we get into Adalat AI, I’d just love to talk about, you know, just the context of the problem, and maybe share a few images from my first few days in a courtroom. 

So the first time I went to a courtroom, this is what I saw. There were towers and towers of paper everywhere. I don’t know how many of you’ve had the misfortune of being in a courtroom in the global south, but this is how most of them look.

Outside these rooms, you have a person standing and their full-time job is to take chits of paper that come as input and spend the next few hours looking for the relevant files. As you can imagine, this is quite painful, takes a bunch of time. In all honesty, this is exactly what Google Search does too. You give it a chit, a query and then it kind of finds you the relevant files. Which I thought was quite interesting the first time I saw this, as an engineer myself.

You go inside courtrooms, you see more paper, rotting paper, there’s this certain smell of rotting paper across courts in the country. Then you go outside the courts, you see people sitting with typewriters and typing documents, not even a keyboard or a computer, just typewriters. Then you go inside courts, you see people working more with paper. There are some computers, but they aren’t quite used. When you look at all of this, it becomes quite clear quite soon that justice in these settings and contexts is not really a question of law, it’s become more of a question of logistics. Then numbers start to make sense. 

This is a New York Times article that says that India has about 50 million pending judicial cases, and at the current rate, it’s going to take us around 300 years to clear this backlog. The situation is quite grave. We have one of the largest under-trial populations in the world. Around 3 out of every 4 people in jail are just waiting for their next date of hearing, languishing in prison. Their family is waiting in courts, trying to figure out the next state of hearing, what’s going on with their families. Make no mistake, behind these numbers is profound human suffering. Right? There’s a very famous saying in the country, which is that the process itself has become the punishment.

What is Adalat AI?

Now, that’s kind of where Adalat AI comes in. We are a legal tech nonprofit that’s building AI solutions for justice systems to make them more productive, so that they can process more cases in a day, and then at scale, you can kind of reduce the backlog of 50 million pending cases. 

Now, I’d love to talk about what exactly do we do, what does this mean? So the big idea is to look at the different parts of a courtroom, figure out the manual clerical pain points, and see how you can leverage technology and AI to either automate them or semi-automate them and assist the process that currently court staff and judges go through. So we’ll walk over a couple of these problems now.

Problem 1: Shortage of Skilled Stenographers

The first big problem is a shortage of skilled stenographers. You see, in a courtroom, nearly every word needs to be written down, and typically what happens is a judge is assigned a stenographer who’s responsible for doing all the writing. But, unfortunately, 90% of India’s courts, and we think that the stat is very similar to other countries in the global south, do not have stenographers.

The reality is that the judge is writing everything by hand in a courtroom. Imagine if I was a lawyer and I asked someone, where were you on the night of 21st April, and they said I was at Stanford Law School, the judge would stop you there, he would write this down slowly while the entire court waits and watches, and then he would ask the lawyer to ask the next question. As you can imagine this becomes a primary bottleneck in the court proceedings moving faster.

Solution: And so our solution here was to build an AI-powered legal transcription solution, which could understand (A) that complex jargon that lawyers love to use, like res judicata, and whatnot. But it could also understand the dialects and accents that change in the country every 100 kilometers or so. And more than that, if that’s not enough, India has many, many languages, and court proceedings in local courts happens in those local languages, and so we’ve built our model to understand legal context in Indian languages as well.

Problem 2: Ad Hoc Processes

The next problem that we’ve seen in a live court is the ad hoc processes. A judge has to see some 50 to 100 courts, depending on the size of the court and the volume on a particular day. Now, when you’re moving between cases, you need to gather the right context for the next case, which involves, as of today, a manual process of collating the right files, figuring out the right pages, and these files can be very, very long. Right? And so, a bunch of time between cases is wasted doing this manual sort of work of gathering context.

Solution: Our solution here was to build what we like to call almost like a Salesforce equivalent, but for courts to manage their cases at one place. So now what happens is we build this tool that can pull your cases for the day, give you the entire context of all the cases, integrated with the AI tool, so that you can do your drafting while you’re reviewing the cases in live court. The idea of this entire flow is if, let’s say, you were taking a certain amount of time, 8 hours to go through 50 cases, in the same amount of time, you should be able to see a lot more cases now. We’ll talk about the quantification of such in a bit. So that’s the second solution that we’ve developed.

Problem 3: Manual Paper Documentation

The third big problem that we’ve seen in a court is all of the manual paper documentation. There’s a lot of paper. In fact, a lot of the 50 million pending cases are not even digitized yet. And so we are making efforts to use LLM-driven scanning and digitization technologies, and we’re building them specifically for legal context so that (A) we can digitize these documents, but also then interpret, understand, and reason on top of these documents, which becomes a major part of the workflow of a justice.

For example, if someone submitted a 1,000-page charge sheet which they need to go over, we build LLM-driven solutions that can help them go through this document in a rather quick manner. And this is not your typical summarization or generative AI. The way that this works is it’s grounded in citations, it’s grounded in highlighting the right parts of a document, given the context of the question.

As you can imagine, as part of building all of these solutions, we have to focus on really fine-tuning models to be able to understand the legal context. We’ve done extensive testing with foundational models, and we’ve found many, many gaps in the current set of solutions, especially when it comes to understanding law.

Problem 4: Litigants Left in the Dark

The next solution that we’ve developed is actually for the litigants. A big problem is that today, litigants are just left in the dark about where their cases are, what’s going on with their cases. They have no idea. Imagine a daily wage worker stuck in a case, he’s trying to understand what’s going on, and someone tells him that his case is currently in an interlocutory order. He would have no idea.

Solution: And so what we are building are WhatsApp-based case updates, case update modules, where any citizen of the country can simply, on a WhatsApp helpline, ask about the details of their case. That information is then first fetched, synthesized, and then delivered to them in a way which would be easy for them to understand. Firstly, in a simplified language, but then also in the Indian language of their choice.

Building for the Judiciary: Security Requirements

I talked a bunch about building solutions for the problems in a courtroom. I want to talk a little bit about what it means to build solutions for the judiciary and security is of the primal need. That’s absolutely the number one thing they care about, and so all of our solutions are built in-house and deployed in servers in India. That’s an absolute no-brainer. They wouldn’t adopt solutions if they’re not deployed in India, and they’re sure that this data is not going to a third party. So we’ve spent a bunch of time hiring the right research talent, ML talent, to firstly build these models, but then also hiring the right kind of engineering talent who can then deploy these models at scale inside of our given country.

The next thing that we’ve also done is we’ve gone ahead and built a state-of-the-art encryption system. What that means is that now all of our data is encrypted at a user level. When you log into the platform, you get a user-level encryption key generated on your computer, and all the data that’s generated on your machine is encrypted using that key, and then sent to the server. 

Which means that no one can access your data without your permission. Even if our servers get compromised, you can’t really make sense of the data because it’s encrypted using a key which no one except for you has access to. 

 I just wanted to make sure that everyone understands that it’s not just about the AI bit, the surrounding context of security is just as important, if not more, and without these things in place, the solution would have never been adopted.

Driving Adoption

And the next part of it is, it’s not just enough to build a solution. We work with justices who can be typically anywhere between 30 to 60 years old, and so driving uptake of this solution itself is a whole second set of problems, and so we’re building an entire partnerships and programming system to train these judges on ground, on using technology to make themselves more productive. This is a large part of our work, too, so apart from technology, we have an entire operations arm with lawyers and state leads that go on ground, work with court staff, work with justices to deploy our solutions, and make sure that they’re adopted.

Progress and Impact

In terms of progress that we’ve made over the last year or so, we’re in about 9 Indian states now, we’re in about 4,000 plus courtrooms in those, which cover 10-15% of the judiciary. What we’ve seen is that courts and judges that use technology like Adalat AI are able to increase their judicial output at a daily and weekly basis by 2 to 3x. What that means is that, let’s say, if a judge was recording, or was able to do 3 witness depositions in a day, now they’re able to do 8, or 6 to 8 witness depositions in a day. Which, you know, at maybe at the judge level, looks like a little bit of an increase, but at scale, when you do this for all the judges in the country, it leads to a case resolution timeline reduction by 30-50%, depending on case stage and time.

We are doing a full-scale RCT, a randomized controlled trial, where we give the solution to a group of judges, we don’t give it to a group of judges, and we evaluate the difference in their productivity. That’s kind of how we are coming up with these numbers.

Global Expansion

Never thought I’d say this, but thanks to colonialism, it’s not just India that has this problem. A lot of the countries in the global south have the same exact legal structure. They use the same legal jargon. In fact, I think this is something I was sharing with Roland as well, that in a lot of African countries, Section 420 is for theft, which is the same in India as well. And so there’s been a lot of consistency in the system, which also means there’s been a lot of consistency in how the system fails. And so now, we think that we can be a global solution for the global gouth. We’re already in talks with countries like Ghana, Nigeria, and Zambia to deploy this solution in courts in Africa as well.

Integration and Mandate

I mentioned that Adalat AI is being used in 20% of courts in the country. I want to talk a little bit about what that means. We are integrated into the live court system, so a judge, every day when they enter their dais on their computer, they turn on Adalat AI, and they use it to record witness depositions, to dictate their orders, to dictate their judgments. We are taught in the judiciary now, as part of the operations training module 2. That’s my co-founder teaching future justices how to use Adalat AI as part of their judicial curriculum to increase their productivity and efficiency inside of a courtroom.

We have also been started to be funded by the judiciary in some states. Finally, in one particular state, which is Kerala, which is on the southern part of India, we were initially mandated in a few district courts as part of a pilot to see how effective these solutions are. More recently, Kerala has gone all-in, and this has kind of broken the news. 

So from 1st November, it will be mandatory across the entire state to use Adalat AI to record witness depositions, and to speed up process of collecting witness depositions. This, in my understanding, is the first time in the world that the judiciary has mandated AI in every courtroom in a particular context of a state. 

Training and Support Program

This is about the latest update, and as part of this update, I want to talk a little bit about the training program that we talked about. We are conducting massive statewide training and support through a training and support plan, which includes daily district-wise trainings. These are lawyers, the ones who are teaching in the screen are lawyers from Adalat AI, who are conducting these training sessions to train judges on the different use cases of how to use Adalat AI. 

We’ve also seen evidences that there are local study groups and peer-led learning groups that are now developing, so some of the people who’ve really become champions of the product have started gathering people into their own peer learning sessions, where they’re teaching their peers how to use the platform, how they are using the platform to be more productive, and we’ve seen largely that these self-led local study groups are extremely effective in driving uptake. Every time this happens, our GPUs start burning just a bit more, and we are a bit happier as well.

We’ve also started specialized statewide training for the court staff, not just justices. For people who do have, or justices who do have court staff, we’re also training them. One thing we’ve realized as part of this training, was initially this was all about how do you use Adalat AI? But we realized there was a deeper problem here. There was a need for digital literacy, there was a need for AI literacy. 

As part of our training module now, we’ve also included things like how to use a computer, how to use a browser, what does it mean to use AI solutions. What are safe ways to use AI? What are unsafe ways to use AI? Why should you not use ChatGPT if you have extremely sensitive data? When is it okay to use ChatGPT? What does it mean for an AI solution to give you an output for you to use it without verification, versus what does it mean to have manual human verification in the process? This is sort of all the innovation that’s going on on the operations side inside of the organization.

We also have weekly office hours that happen remotely on Zoom to be able to be a lot more hands-on with respect to the support we provide the judges. Apart from that, we’ve set up an entire customer feedback pipeline and a success team that constantly helps us monitor the platform’s performance, any bugs that come up, and sort of subsequent fixes. 

We also have built dedicated WhatsApp community. India runs on WhatsApp. Slack channels wouldn’t work. We build these communities where judges come together, share their problems, and then we kind of work with them to solve their issues, and that also becomes a very natural way for us to improve our product.

The Team

I want to talk a bit about the team that’s doing this. We are about 30 people. I’m actually the AI in the company. My co-founder is the Adalat. Adalat is an Urdu word that means court, so he’s the lawyer in the team. He practiced in the Supreme Court for about a decade, did his law from Howard Law, studied econ from Harvard Kennedy and Oxford. 

Really saw the problem firsthand as a practicing lawyer in courts. Then decided to do something, or take a systemic approach, and I left my for-profit startup journey to join this mission to provide timely justice for all. On that note, we are functioning as a non-profit entity, for a various number of reasons, but primarily because it’s very hard to get budgetary allocation for technology in courts right now, because there is no precedent for it. In fact, judges in District Court use Ubuntu and LibreOffice, because they don’t have the budgets for Microsoft licenses or Google licenses.

Vision

The goal that we have is that a lot of digital public infrastructure in India, like the Aadhaar, which is our identity sort of stack, and the UPI, which is our finance stack, were built outside the system and then taken into the government to make building the solution more effective, and then deployment is managed by the government. We are doing something similar, but for the judiciary. We’re building this outside the system with a group of highly talented individuals, and then taking it into the system.

That’s a little bit about Adalat AI. I would love to hear thoughts, questions, that might come up.

Q&A Session

Roland Vogl: Great. Well, thank you so much, Arghya. This is… this is a great presentation. It gives us a really good sense for what you’re up to, and it’s truly exciting. So, I know Bruce has his hand up, but Marzieh has asked a question before already. So, Marzieh’s asking, how do you deal with the fact that LLMs don’t understand the law, per se.

Arghya: Am I correct in my understanding? Is the question, how do you deal with the fact that common man doesn’t understand the law?

Marzieh Nabi: No, the LLMs.

Arghya: Okay, LLMs don’t understand the law. Yeah, you know, initially we started with benchmarking exercises. I think what we did was we took a bunch of foundational models, we really focused on building datasets that represent the different tasks, so I can talk about the tasks. 

One is, let’s say, legal transcription. You have a bunch of audios that were recorded in court by actual justices, and then you have the right transcript for it. You kind of run your AI models through them, you figure out what the baseline performance is. We sort of realized that there were many gaps. So, one-word answer would be data, but the more elaborate answer would be that there was an extensive effort made to scrape data off the internet with respect to law in the Indian context, and then we hired lawyers in-house, who then annotated the data, helped us build expert datasets, which our models were then trained on.

Once these models were trained, we did blinded evaluations on the output of these models with lawyers. Lawyers were given the output of, let’s say, model A versus Model B versus human, and they were asked to choose which output do they prefer the most. We’ve reached a stage now where Adalat AI models’ outputs are preferred 70-80% of the time over all other models, including the largest models, like let’s say ChatGPT.

Roland Vogl: No, no, I was just saying, so your question goes to the point that you can own… I mean, you know, it’s the LLMs, they predict the next token, right? And so you can train them to get fairly accurate, but to have 100% reliable answers is not possible, at least not at this point in time, but the point is that you have all these use cases, I guess, where you’re not really… it’s not really about legal reasoning and applying law to fact and needing a 100% reliable solution, answer. It’s more about, like, streamlining digital processes, transcripts, and so on. Right? And then you also… and in the areas where you have… actually, where you’re applying the law to facts, then you’re… you’re using LLMs as opposed to, like, deterministic systems, but you have lawyers kind of help you, improve the outputs, the reliability.

Arghya: I elaborate on that a little more. So, basically, the way that we started was we picked the simplest task. Speech-to-text, no risk of hallucination. You give an audio input, it gives you a text output. There could be mistakes because of the way that you are saying certain words. We really started with tasks which do not require a lot of reasoning prowess from these models. And we are slowly working our way upwards with adoption of the other features that require, let’s say, more LLM reasoning prowess. 

For example, given a lengthy charge sheet, and you ask a question, can it find the right places in the document, and then answer those. Those are much more advanced use cases. We actually started with very simple use cases where the ML solutions have been proven to be timeless and have been working for decades now.

This also helped us with adoption, because imagine going to a court and telling them, we’re bringing you predictive technology that’s going to do your job, that’s really going to not fare well. And so, this was a process of continuous iteration and building trust with the audience, too.

Marzieh Nabi: I just asked one clarification, so I know we are staying on this for a bit, but, like, the expert data that you were talking about, the lawyers that you hired to tailor the data and all of that, how did you incorporate those into the LLMs? Did you incorporate, or you just use them for the evals purposes?

Arghya: No, no, we do fine-tune our models using in-house. 

Marzieh Nabi: Okay. Using reinforcement learning?

Arghya: That would depend on a task-to-task basis, essentially. So some tasks might require reinforcement learning, where the model is… it’s not enough for the model to just see what the right output is, but it’s almost important for the model to practice these problems and come up with the right answer on their own.

Marzieh Nabi: That makes sense.

Arghya: The right way that we think about this is, you know, learning is a three-part journey. If you think about it, you first read textbooks, then you see solved examples in textbooks, and then you do a bunch of practice problems that are there at the end of the textbook, where you don’t quite have the entire solution, but you have the final solution. 

We take our models through the same journey. We first have them read a lot of textbooks. And then through supervised fine-tuning, we show them what it means to have a problem, and then solve the problem in the correct way and arrive at the correct solution. Then we give them practice problems where we use reinforcement learning to have them sort of solve these problems on their own and figure out the right way to do it.

Marzieh Nabi: Sounds good, thank you.

Bruce Cahan: This is great. I’m really excited to see this activity. We have started an Access to Justice Alliance for All programs, and so I’d love for you to be part of that. We also at Stanford have something called the Center for Human Rights and International Justice, and the head of that, David Cohen, has been, for many years, helping the judges in Indonesia understand better how to organize their justice system, and I’d love to make that connection, so maybe I can have your email, and I can do that.

Arghya: Totally.

Bruce Cahan: And then substantively, and Roland touched on this, I think we have a taxonomy called Folio that in part was developed at Stanford by Margaret Hagen and her group that, in effect, is almost like the biology of the law. You know, the principles of tort, or the principles of theft, neighborhood, etc. I do wonder, particularly in the context of our working group, how there are certain universal taxonomic features to our different justice systems that all require proofs at a different level of particular elements of a crime or of a cause of action. Training to understand that would seem to be a little more universally relevant here than predicting the next set of words in a judicial decision.

Arghya: Absolutely. I think, firstly, thank you for the help. I’m definitely going to take you up on it, and I’ll reach out to you. With respect to your idea about a universal taxonomy, I completely agree to that. In fact, we’ve been trying to figure out what are the best ways to help these models learn law, which is also part of the reason why I was at the law school. I wanted to meet Roland and Robert, because I was there last year for the UN AI for Good panel, and I had a great time. I got to learn about the philosophy of the CodeX group, and I think there’s a lot of common ground of research between what we do, and what CodeX wants to do, and what CodeX aims to do. That was kind of the purpose, and so for all the folks here, if there’s anyone who’s interested in pursuing research in these directions, we’re extremely nerdy, we love publishing, and we’d love to collaborate on, you know, getting something out together.

Roland Vogl: I’m gonna try to go through some of those questions quickly. Sushmita is asking, how you deal with language challenges in trial courts, because you, you did mention that early on, maybe Sushmita didn’t hear that, that you, you, your system works in all the different languages spoken in India, right, and the courts that are running their proceedings in those languages.

Arghya: Yeah.

Roland Vogl: Running their proceedings in those languages.

Arghya: Yeah, I can… I can take that. It’s a very hard problem, just by the way, because it’s not only just the number of languages, the number of dialects of every language are so different, and also, even with English, the way that a judge, let’s say, in the southern part of India talks in English is very different from the way that a judge in, let’s say, the northern part of India works with English. 

Again, at the risk of being redundant, data is king. We really had to go through a massive effort of data collection, and so I talked about lawyers who help us build these datasets, these expert datasets. We also have language experts who come in, work with lawyers to help us build datasets in multiple languages in the context of law. Right? Legal English is so different from normal English. You have, I think, an influence from Latin. 

The problem becomes much more graver when you move to an Indian language. For example, I could understand Hindi, but there is no way that I’m understanding how legal Hindi sounds, right? Really, the effort there was in using a lot of human effort to build these datasets.

We literally have people read out judgments in the way that judges do. We hire experts to do that, and then we kind of use that data to train the models. All of this is also because we can’t use the data that’s generated on our platform to train our models because of the sensitive nature of the data, until someone actually gives us explicit consent. Hopefully that answers a part of your question.

Roland Vogl: So, Abbeit was asking if you had any opinions on the current limitations of open source versus commercial LLMs, and ability to understand and analyze legal context.

Arghya: Absolutely. As I was telling you, Roland, in fact, we recently published a paper on this, so the effort was to take all of the public exams that are present in India that gatekeep the profession of law and let humans into it. So, for example, the CLAT exam for entry into a law school, the CLAT PG exam for entry into a law school for the master’s program, the judicial services exam, the higher judicial services exam, and we ran a bunch of these foundational models, open source, closed source, through them, both objective and subjective exams. With subjective exams, we did a double- blinded evaluation where we sort of gave a human paper and an AI paper to the same person who is a vetted person who can evaluate these papers, and we sort of did an entire comparative analysis.

There is a big gap between open source and closed source models. If you’re interested, I’m very happy to share the paper afterwards with the group as well. But there is a huge gap between open source and closed source models, but TLDR is with objective papers, these models absolutely nail it. 

They beat the human topper every year by a margin, but when it comes to the subjective papers and writing subjective answers to critical thinking questions and questions of drafting, these models don’t do great at all. In fact, the human evaluators seem to be punishing the model outputs a lot more than they do humans, and we’ve kind of done an entire work on this. 

So I’m not sure if it answers the question, but the gap is significant. Even with the best foundational models, I think the gap with humans right now for subjective exams is significant, which says something about the profession of law being more of a human-human thing, as opposed to, you know, just pushing… putting a machine in the loop.

Roland vogl: Okay. So, great, thank you. So, maybe just make two very quick questions, because we’re a little bit over time, but I know, Yogendra has a question, and you put it in the chat, maybe you just want… and you had your hand up, maybe you, and then… and then Kathy, and then…

Yogendra Jain: So, my question is very simple. Basically, in northern part of India, we have a lot of things in revenue codes and in civil codes, handwritten things we use. Right, so how effective is your OCR system? What systems are you using? And in the coming time, deploying them in comparison to the current system, where we do not have any revenue going into it? How much effective would it be, in your understanding? How much time would it take?

Arghya: Yeah, the unfortunate reality is that status quo is actually to take those documents and type them out. So that’s status quo process, right? So you do any sort of digitization, it’s a lot more effective. 

Now, coming to your question about what kind of models do we use. As of right now, we don’t support handwritten documents, we support only printed documents, and that’s fairly doable with a large, sort of family of models, you know, starting from Tesseract OCR to everything like these multimodal, smaller models that we have available now. So, really, OCR is… our digitization record is more of a scale problem, an ops problem. How do you actually take all these papers and click images of it, as compared to a ML challenge right now. I think the main ML challenge, or the tech challenge, lies in how do you digitize handwritten documents, and that’s something we’re still working on. It’s an active problem that we are working on.

Roland Vogl: All right, so, so maybe Kathy can ask the last question, and Arghya, I think if people have more questions, after this, after this call, what’s a good way to reach you? Maybe, maybe you can put your contact in…

Arghya: Yeah, I’m available on, my name at adalat.ai. I can also share quickly my LinkedIn, so that people can reach out. Very happy to chat about this. This is something that we do day in and day out,

Catherine Atkin: Just very quickly, I want to say this is such important and exciting work. I mean, this is really breakthrough. Creating this opportunity for justice at that scale is so commendable, so thank you. I was just wondering about the business model. You mentioned that you’re a non-profit, and how do you think over time, as there’s a drive for more functionality, do you see private businesses coming in to serve a purpose? And lastly, are there other regions in other parts of the world—is there a community of people doing this kind of work?

Arghya:  Great question. This is something we actively think about all the time. Even though we are a non-profit, we need a model to sustain ourselves eventually. There are two or three ways in which this could play out.

One is that we’ve already started seeing that the other set of people who come to courts, apart from justices and citizens, are law firms and lawyers. They come in, see Adalat AI running in courtrooms, reach out to us, and say, “Hey, this looks very interesting. We have the same problems. Can you provide us with access to your platform?” One of the ways in which we think about being able to sustain ourselves is cross-subsidizing the impact by selling to the private space. We’ve already started repackaging our solutions for the specific problems of law firms and also arbitration centers, which are almost like private courts. So that’s avenue one.

The second is, when we started out, there were no budgetary allocations for technology in courts, but that landscape is changing fast. We anticipate that right now we have one state which has come forward in funding us, but once this solution really picks up and people see the value—and we are able to prove that if you use this technology at scale, you can reduce the case resolution timelines by 50%—we are very optimistic that we’d be a great candidate for government adoption. 

So one possible exit, if I may, would be that the government consumes the systems, runs this at scale, and becomes the payer and doer at scale.

The third is, turns out India’s not the only one with these problems. A lot of countries in the global south, even a lot of wealthy countries, have the same problems. So we are now considering if the same set of solutions can be sold to some of these other countries that are doing well economically to sustain the impact on the global south.

Catherine Atkin: Thank you.

Roland Vogl: Alright, thank you so much. We’re 10 minutes over time, but this is great, and we really appreciate you joining us, Arghya, so late at night over in India. It’s amazing work that you’re doing, and I think you hear from this group how much we’re cheering for you, so please keep us posted. Thank you so much for joining us. I hope you have a good rest of your night. Thank you, everyone, for all your great questions. Big round of applause for you. And please stay in touch. I’ll see everyone at the next group meeting.