For many urban renters in India, getting a security deposit back can become a drawn-out battle. Deposits of ₹2–3 lakh are common in cities such as Bengaluru, yet tenants often lack clarity on what deductions are legitimate, which laws apply to them and what legal recourse is available when a landlord refuses to return the money.
For Siddharth Agrawal, a Bengaluru-based product manager, the problem became personal. After his landlord refused to return his security deposit, Agrawal was told: "eJagriti mein bhi kuch nahi hoga." His response was to build a product over a weekend.
Called Recourse, the AI-powered tool helps tenants generate pre-litigation legal notices based on the laws applicable to their state. Rather than acting as a generic chatbot, it uses a state-specific retrieval architecture to identify relevant legislation before an LLM drafts the notice.
Users answer seven questions about their dispute to get what Agrawal calls a filing-ready pre-litigation notice, grounded in texts including the Karnataka Rent Act, Model Tenancy Act, Consumer Protection Act 2019, Section 74 of the Indian Contract Act and Section 80 of the CPC.
The approach reflects a larger challenge in building AI products for India's legal system which is the law itself is fragmented.
Why rental law is difficult to build for
Tenancy is a state subject in India, so rental rules vary significantly across jurisdictions. The Model Tenancy Act, 2021 tried to standardise the framework, but adoption is voluntary and uneven, as of mid-2026, only a handful of states, including Assam, Andhra Pradesh, Tamil Nadu and Uttar Pradesh, have substantially aligned with it.
Most others still run on older, tenant-protective legislation, some dating to the 1940s, or sit in a transitional phase. A tenant moving between cities can face an entirely different legal regime on notice periods, deposit caps and eviction grounds.
This cuts both ways including old Rent Control laws making eviction a years-long court process even for non-payment, discouraging landlords from maintaining properties earning below-market rents. Deposits are a flashpoint too as Mumbai and Bengaluru landlords commonly demand six to ten months' rent versus the two-month cap under the Model Tenancy Act, with vague "wear and tear" clauses fuelling disputes. Many rentals are also informal, or structured as 11-month "leave and license" agreements to dodge registration under the Registration Act, 1908, which weakens their standing as court evidence.
"The challenge is not simply that tenants do not know their rights. The legal framework itself can be difficult to navigate," Salman Waris, technology lawyer and founder of TechLegis tells AIM.
For an AI developer, this is a fundamental technical problem. A general-purpose LLM can write a convincing notice without identifying the correct legislation.
Waris also flags the risk of outdated references and hallucination as Indian law changes. The 2024 shift from the Indian Penal Code to the Bharatiya Nyaya Sanhita is one example. Agrawal's solution is therefore built around retrieval rather than generation alone.
Why Recourse doesn't simply ask an LLM to write the notice
The architecture has three layers including retrieval-augmented generation that searches state-specific legislation, an AI layer that filters what's retrieved, and an LLM that drafts the final document.
"There is a RAG model which is retrieving the data. Then there is a foundation layer where the AI is filtering out the data. And third is the LLM, which is actually drafting that process," Agrawal explains to AIM. He's currently using Redis and a vector database for retrieval, with Gemini 2.5 as the LLM, while evaluating GPT-5.5 and Sonnet.
The product is built on ReactJS and Convex, with Claude Code used during development. The retrieval layer matters because the system must establish which law applies before generating the document — "Karnataka Rental Act, Maharashtra Rental Act, Bengal Rental Act, it will go there," he says. Without that grounding, an LLM "will hallucinate in just a second."
There's a distinction between an AI generating a notice and one drafted or certified by an advocate. Under Indian law, no requirement mandates that a legal notice be prepared by a lawyer, a self-drafted notice is valid if factual, clear and formally worded, and AI tools aren't explicitly barred from drafting one either.
But Waris points to the credibility gap this leaves. A notice on an advocate's letterhead, with a Bar Council enrolment number, carries more weight than a self-drafted one, something an AI system can't authentically produce.
For procedurally demanding notices, such as Section 138 of the Negotiable Instruments Act (cheque bounce) or Section 80 CPC (suits against the government), strict formal rules apply, and an LLM that cites an outdated law or misses a mandatory element could render the notice defective.
Waris flags three further risks stating courts favour notices sent by Registered Post or Speed Post over electronic ones; responsibility for any error in an AI-drafted notice rests entirely with the sender, since an LLM can't be held professionally accountable; and AI tools may miss limitation-period nuances a lawyer would catch, making professional drafting the safer route for high-value or complex disputes.
That's why Agrawal doesn't see AI as replacing lawyers in Recourse, instead, lawyers are becoming part of the product architecture.
Agrawal's next step is onboarding legal advisors to verify notices and improve the system. "Right now the accuracy might be 80%. How can I trust it?" he asks — a hallucinated answer in a casual chatbot is inconvenient, but an incorrect legal provision in a notice can materially affect a dispute.
Lawyers would play two roles which is human verification (eventually leading to an advocate-certified notice) and training the model on real-world scenarios. "I want legal advisors to come and train the model. That is very important," Agrawal says. This maps onto Waris's concerns: verification closes the credibility and procedural-accuracy gap, while certification gives users a document with the letterhead and evidentiary standing an AI draft lacks alone.
Agrawal wants Recourse to eventually analyse rental agreements and warn tenants of problematic clauses before they sign — "it should actually tell the user, hey, this is a problem in your rental agreement." He plans to make uploading a rental agreement mandatory before drafting a notice, pushing the product from reactive dispute resolution toward preventing the ambiguity around deposits, rent hikes and maintenance that persists in informal, unregistered agreements.
Early traction, but a long way to go
When Recourse went viral, more than 3,000 people visited the platform, over 1,000 registered, and roughly 1,500 which is a 51% completion rate, finished the process. The project also picked up 56 GitHub stars.
The product works better in Karnataka than Delhi.Karnataka users can complete it in under five minutes, while in Delhi "the rental app does not work... we are not able to fetch it," Agrawal says, adding he's working with legal advisors on a fix.
Recourse remains a small product born from a personal dispute. But its real technical challenge is bigger: navigating India's fragmented legal landscape,archaic Rent Control Acts, an unevenly adopted Model Tenancy Act, largely informal agreements, without turning legal uncertainty into automated misinformation.
The test isn't whether an LLM can write a legal notice, but whether retrieval, jurisdiction-specific grounding and human oversight make it accurate, procedurally sound and credible enough for a tenant to act on.
Your reaction
Discussion
What's your take on this story?
No comments yet — be the first to share your view.