You're absolutely spot on on most of these points and have clearly done your homework. Nevertheless, your skillset and experience will certainly be in huge demand and many organisations and industries would be lucky to have you, or you will probably identify an interesting niche to do something interesting in. All the best, you'll clearly be just fine (even if AI will end up making us all irrelevant one day, but probably a little while after it bursts like a wet bubble :)

I am working on a guide it is going to take me some time, and end sems are also approaching so I cannot work on it for the time being, but I will make a super detailed post by 10th of June explaining how everything works and go into every nitty gritty of it. Also, as I said, I am trying to make my database public.

Now, commercialisation: it is a crapshot made using a sling owned by someone else.

The problem is that the market is flooded with AI wrappers doing the same thing, connecting the Legally India API, Claude API, or whatever cheap AI model they could find or got credits for from the company at the time. Now, if I do the same, I will be offering a subscription for ₹1500-3000 a month, but then the output quality will degrade so much that it will become useless for everyone and anyone using it, no one will be using it, plus I will still be making nothing but losses as the margins are very thin at this price point. It will be the 100th AI legaltech company, claiming to revolutionize the field.

To do something truly good, Harvey AI type, I will have to incorporate a company, hire people, and regular run of the mill engineers aren't going to cut it. I am talking about top-tier talent, SDE 3 and above. These guys are in extremely high demand, so they'll charge even more to enter something this un-proven with someone who doesn't have a background in tech, if at all. At the top, for building everything, what I did was essentially make a tent. To make an enterprise product, you're looking at a building made of concrete that follows western building code. Handling 20-30k concurrent enterprise users, load balancing, cyber security at that scale is not just beyond my expertise level, but my comprehension as well. Now the building part involves a frontend and a very robust backend that isn't held together by duct tape and my hopes, and actual contracts with data providers. On top of that, I will have to make a database of my own that will have to get much bigger, deeper and nearly error-free, and would require constant updating. Think of daily updates from every High Court and District Court, making my own headnotes, summaries and stuff like that. EC2/Lambda instances to run the whole thing, and for AI models, Google Vertex or AWS Bedrock, with region/cross-region inference due to data residency requirements. This, by the way, was the least of my problems, as it is manageable.

The bigger worry is that I will have to subscribe to either the reserved tier or Provisioned Throughput tier. What both of these mean is guaranteed compute capacity no matter the load, which sounds good, but there is a very huge but: I will have to pay for the capacity even if anyone is not using it, I will be paying the same amount regardless. And you cannot use demand-based tier because those institutions who are paying good money are not going to tolerate downtime because your cloud provider ran out of compute, which happens often enough that you need to be plan for it.

I had an idea where I thought about using audio streams from virtual court meetings, converting them to text, linking them with the cause list and making a near-instantaneous court status tracker and data collector. But that would have required permissions from courts, it would have been prohibitively expensive and prone to errors as well, though it would have led to a great database.

Harvey AI is also facing similar issues in countries where they don't have a lot of customers: hardware sitting idle in data centres, still paying for it, making losses. But they charge $1,200 per seat per month with a minimum of 20 seats, so that comes to roughly $288,000 annually, and you can somehow make it work at that price point. It also needs to be taken into consideration that they raised over a billion dollars and are still burning money. Their token cost is higher as well due to the exact same reason I mentioned above: dormant compute.

For compliance, in order to sell to any decent organisation internationally you will need GDPR compliance, DPDP for India, ISO, and last but not the least, SOC Type II. SOC Type II deserves some extra focus: it is very important, costly and time consuming. To put things into perspective, this is the kind of certification American companies spend thousands of dollars on, and an entire company was set up in the US that claimed to use AI in the process to hasten it, called Delve.co . These guys are doing nothing but fraud by forging reports, and start-ups knew and were complicit. This should explain how time consuming SOC Type II is.

Now, even if I somehow secured funding and cleared all these hoops before going bankrupt, this business model is unsustainable and very short-lived at best. These companies are entirely reliant on other companies for the most important part, that is intelligence. You can make servers with enough money, but you cannot spring up a model out of thin air. I know someone will say they can use an open weight model, but the problem is not just that you are banking on Chinese companies continuing to release them it is that you only need one geopolitical event, one export control regime, or one strategic pivot for that dependency to collapse overnight. The probability is not even the point. The consequence of being wrong with no fallback is. And as of now closed models are still better.

The free cash flow of hyperscalers is coming to an end, and as that comes down, the funding for the labs will also start to dry up. But it is actually worse than that. Hyperscalers are simultaneously bleeding FCF on capex and not getting the physical capacity they are paying for on schedule. Cancellations have quadrupled, the total capacity under construction has declined for the first time since 2020, and transformers and electrical equipment sourced from China are holding up billion-dollar projects by a year or more. The compute supply that labs depend on is more constrained than the headline spending numbers suggest, and the return on that capex keeps getting pushed out. You cannot keep up the pace for very long. Then, when they get hungry for more money, because you cannot just keep making losses when your funding pipeline has dried up, the first entities they will go after are these middlemen. Not because they are small consumers, but because their entire business depends on them, and that is exactly why they have no leverage.

Think of movie distributors and how they twist the arms of cinema chains. They know PVR is sitting with around 1,750 screens that need to play something on them. They know they are desperate, so they do that, even though film distributors need them too. Customers are also not dying to go to PVR because of OTT and Telegram, the alternatives don't have to be perfect they just have to be cheap and good, so they are getting hammered from the supply and demand side. I know institutional contracts don't work like that, but I believe this analogy gets what I am trying to get through. The problem is that once these companies start to get squeezed, they can either absorb it or offload it to the customer and risk them migrating to their in-house tool, and if more customers leave, your costs increase, thus creating this deadly cycle. They aren't able to turn a profit even now; the future looks bleak.

Another problem is that you're nothing but a middleman, and if anything the internet has taught us, it is that the middleman gets crushed sooner or later. To be more precise about it the specific thing that is lethal here is the narrowing capability gap. As frontier models get better at legal reasoning out of the box, and vibe-coding the delta between using Claude directly and using a wrapper built on top of it shrinks toward zero. Right now every company is focusing on coding benchmarks and making the best IDEs and CLI tools, but the day they conquer that market they will come for this one with all their engineering experience and might. Plus, as models are getting more capable, their ability to sell simple solutions is short-lived at best. Think of what happened to CRUD developers: you cannot bank on complexity being the only thing keeping you alive.

The only thing that can save them in the long term is government protection, which in my opinion is very unlikely, as these guys aren't necessarily serving a purpose. Though I can be wrong as well. The thing that can truly save them is proprietary data and in my opinion this is the only thing that can save them. By data I mean a corpus of cases with their own headnotes, books, reports, articles and stuff like that, and not having to rely on third-party services for data. The issues are so big that you cannot just solve them. Palantir benefits from AI because of how good their management software is and how deeply it is integrated everywhere. With AI wrappers, I don't think the same can be said. I am making a very crude prediction so take it with a grain of salt they have a life of 10-15 years in my opinion.

Lastly, and this is the least important point: I feel it is very wrong to charge for plumbing that someone can do themselves with some learning. But it is not very important, as I am not really faced with the choice here, so saying how I would have acted when presented with the opportunity is rather abstract. What I can do, and will do, is share what I have built.

I'm a 2nd year law student, not a professional. This workflow/system is the result of my laziness and my lack of ability to trawl through material to complete assignments. For the first 3 semesters I was just downloading research papers and feeding them to Claude Sonnet with my writing sample. Output was decent for assignments, but I knew this isn't it, citations can be better, more depth can be covered.

This semester I had a mandatory graded moot and I used SCC Online, AIR and their likes for the first time, and the sheer inefficiency boggled me. This clearly wasn't going to work in the long term, I have 6 semesters left, which comes to around 36 assignments, and I need better. So I started with API reverse engineering. Not a huge problem for me, I have done this in the past, I have worked with documented trading APIs and I have been messing with websites for 3 years now, so I had a very decent working knowledge, knew how stuff works. It isn't particularly difficult either, all you need is a sense of architecture and curiosity; I'm a law student who came from a commerce background with a fascination for computers, so it's not difficult. I kept building on top of that SCC Online MCP, and now the problem is that there's so much to download and index that it is becoming difficult. Even setting up a cron job at this scale is proving to be difficult.

I am actively trying to make the offline corpus bigger, I aim to include as many reports as possible, Gazettes, Parliamentary Debates, laws that were drafted but never passed, articles from public journals in my DB. Right now it just has laws and a few hundred important and major reports. The problem is the majority of the stuff is not organised, so downloading, indexing and organising is a huge task. I hit a wall multiple times and still hit one with every new source, but it is well worth it. I use this solely for assignments, as I was never interested in moots and research papers, so I am not hurrying like my career depends on it.

That said, I am trying to fix all the errors and make the DB public, access to structured data shouldn't be a bottleneck for anyone. I cannot make the MCPs public for obvious reasons. It took me about 2 weeks with a lot of work, around half a dozen all-nighters and once I worked 8 hours straight to put things in perspective, and that's with my background; the part that took time was trying to come up with ways to improve it. From scratch with no background, 2 to 4 hours a week, it should take anyone willing 5 to 6 months. It is a little difficult to make and work with, but it is a complete game-changer for anyone willing to put in the work. Any small law firm or solo lawyer can and should be able to adopt this as well once it is more stable and bigger.

I'm also working on a model interconnect so ChatGPT, Claude, Gemini, Kimi and Deepseek can debate each other and find flaws in the output. This would strengthen the reasoning capability significantly, and the best part is it would still be cheap af, though it is proving more difficult than I thought.

This is also better than commercial tools because it is not bound by a lot of the institutional stuff they are bound by, liability hedging, locked-in models, fixed pipelines. They also cannot just reverse-engineer APIs, download stuff off the internet and call it a day, and they have to build something sustainable as a business. Since you're paying for the subscription of AI services directly, you get far better value, as the API itself is much costlier, and after the overhead of the company you're either paying very high prices or paying for it in the quality.

The best part about this entire system is that you're always in control of everything, the reasoning capacity, data, model type, costs, the offline corpus, how you want the output to look like. The entire thing is practice-agnostic, you can fetch data, make a database and start it. Plus the custom instructions in CLI are followed much more thoroughly than they are in the web. It is cheap as well, especially if you use opencode go.

If anyone wants to build something like this and needs technical help, feel free to reach out at LIunofficialtechhelp.contact@proton.me. Do not ask me for the MCPs, I won't be sharing them for obvious reasons, and don't try to buy them either, you'll be wasting your time. I can guide you through building your own, but sharing mine crosses a line I'm not willing to cross. I'll be lurking on the thread for a few days as well if anyone has doubts.

Here's what I've been running-

I reverse-engineered the APIs of SCC Online, Westlaw Classic, and Westlaw Asia, and made MCPs out of them. You need a subscription, as it's based on cookie-based authentication replay, but it works. You're kind of violating ToS, so don't abuse it, or else those guys will abuse you, and you don't want that. Don't try to scrape any of the closed DBs either. Thomson Reuters went rabid when ROSS Intelligence tried to pull it off; they had VC backing and still got crushed. You don't stand a chance. The MCP is also significantly faster than the actual SCC Online website; search and fetch take a fraction of the time, just to give you a sense of how well it performs. I use Playwright instances for mass-scale downloading of data and cookie capture. The Playwright instance has a keep-alive function; it simply keeps visiting pages every few minutes, and once a cookie is captured it stays valid for a good amount of time. Make sure to enforce proper rate limits, or you'll nuke your personal or institutional subscription.

You can do the same with AIR Online and The Hindu. I tested AIR Online with curl and direct API requests and it worked, but I didn't find much value in it, so I didn't bother adding it. I didn't try The Hindu, but the console does reveal an API key, so it should be theoretically possible. Added an Indian Kanoon MCP using their official API. Plus, there's a very popular library whose MCP you can use by donating to them, so it fetches books as well.

For the offline corpus, I downloaded every piece of legislation and report I could find, then rented some GPUs from E2E Networks (they're India-based and the rates are reasonable; note that you'll need to complete KYC on E2E before you can spin up compute). Use the PyTorch environment. It's the cleanest base to build the OCR pipeline on. Ran OCR first with Tesseract, which is very fast, and then for the files that had bad recognition I used Chandra LLM OCR, though its speed varies a lot (pro tip: make sure you're using vLLM and not HF). An L4 is sufficient for Tesseract, but for Chandra you'll want an A100. Don't waste time trying to run Chandra or Surya OCR on E2E's L4 instances; they have serious driver issues.

Structured everything into an FTS5 SQLite DB (FTS5 because it gives you better search capabilities), made an MCP out of it, and connected all of them to Claude Code CLI and opencode. There's a hierarchy in how the systems are queried: one is checked first, and if that doesn't satisfy the query, the next one is hit, and so on. The best part of this setup is that the cases are saved in .txt format, so the AI can actually read them, unlike what these sites do by serving non-machine-readable PDFs.

Modern frontier models can either handle 60-70 agents at once or, failing that, give you a 1-million-token context window. For instance, Kimi K2.6 can launch up to 300 agents but caps at a 256k token window, whereas Deepseek v4 Pro offers the full 1M context and can also launch multiple sub-parallel agents, just not as many as Kimi. Either way, you can research books, cases, and reports all at once with zero issues. That said, having 1M context doesn't mean you should fill it up; retrieval quality degrades as the window fills, so treat it as headroom, not a target. Also, don't try any parallel agent tomfoolery with Claude; you'll burn through your limit in under three minutes. Once the cases and material are saved, one agent is deployed per file, each saving its output in a separate folder in Markdown format, and finally one agent synthesizes everything. Opus 4.7 is the best pick for the final synthesis: it's the smartest model out there, and it'll consume very few tokens because the cheaper agents will have already chewed everything down and added the digestive aids. A final agent then checks the synthesis. The possibility of errors slipping through is low but real, which is exactly why humans still exist in the loop. Use Obsidian for reading the Markdown output, and make sure the Markdown is formatted properly so it converts cleanly to .docx via Pandoc.

On tooling, I suggest using opencode CLI, as you get access to opencode go, which is quite cheap. Claude Code eats tokens like anything; to do anything meaningful you'll need the $100/12000~ Max 5x plan, but ₹2000 pro plan will work if used cautiously. Codex CLI is also quite good with ChatGPT Plus. The usage limits are pretty generous even after they constrained them, especially compared to Claude. This workflow works best with CLI tools, not the apps of these platforms; I haven't tried the apps much, as it's just easier to work with these in the CLI. Also, adjust the reasoning capacity of the models based on the complexity of the work. Running Opus 4.7 on max reasoning will burn through your quota in minutes. Make sure you're using a custom directory and have custom instructions, such as telling the model to rely only on the supplied material, produce a review, and give the exact source.

A round of research can take anywhere from 30 minutes to 2 hours, depending on how many issues crop up, since the whole thing is still held together by duct tape, hopes, and wishes; that said, most issues resolve themselves in minutes. As for build time, I had a working knowledge of APIs and system architecture, which sped things up manifold. AI in the CLI also helps a lot with the build itself: even if you just hand it a HAR file, it'll reconstruct the API payloads for you. Took some doing, but scarce material and slogging through archaic portals stopped being the bottleneck.

Whereas I think I’m a decent legal researcher as a second-year law student, how can I strengthen my skills further? I’ve already worked on most of the generic things familiarity with SCC and Manupatra, reading judgments, etc. Most of the time, I do end up finding something useful, but it takes me considerably longer to find it. How can I improve and become more efficient at legal research?