CounterDeal: A Local-First Neuro-Symbolic AI Real Estate Terminal

Title

CounterDeal: A Local-First Neuro-Symbolic AI Real Estate Terminal

Meta Description

CounterDeal is a local-first neuro-symbolic AI real estate terminal that uses gemma2:2b for language, TypeScript code for MahaRERA compliance, and TabPFN to calculate deal probabilities. In a Pune brokerage pilot, it boosted conversion by 40% and cut compliance checks from 45 minutes to 3 seconds.

Article Content

The real estate brokerage industry has long faced a dilemma: the financial records and property transaction data of high-net-worth clients are highly sensitive, and handing them directly to cloud-based AI tools means risking data leaks. Yet letting a large language model handle compliance calculations on its own is almost certain to produce hallucinations—and in the face of strict regulations like MahaRERA, a single fabricated number could trigger a legal dispute. Developer smtherer built CounterDeal for the real pain points of real estate brokers in Pune, India: a fully local, neuro-symbolic workbench that separates the language model from deterministic code.

Core Architecture: Let the Language Model Do Language, Let Code Do Math

CounterDeal's design philosophy isn't complicated, but it hits the mark: split tasks by nature—natural language understanding goes to a small model, while regulatory calculations and probabilistic inference go to deterministic code or specialized tabular models. This "division of labor" directly avoids the two most common sources of accidents—data leakage and numerical hallucinations.

Local AI Runtime: gemma2:2b + Ollama

The workbench's navigation and the "Buyer Concierge" feature are powered by gemma2:2b, running locally through Ollama. The model is small in scale but sufficient for intent recognition and conversational retrieval. Crucially, all client requirements and financial queries stay on the local machine and never pass through any third-party server. For brokerages handling high-net-worth client data, this value far exceeds the model's capabilities themselves.

Ollama is an open-source local LLM runtime framework that packages model weight downloads and inference services into simple command-line and API interfaces, supporting macOS, Linux, and Windows. Developers only need to run ollama run gemma2:2b to start the model on consumer-grade hardware and call it through a local HTTP interface (default port 11434). Unlike cloud APIs, all inference in Ollama happens on the local machine, and requests never leave the device—making it a common choice for privacy-sensitive scenarios.

Gemma 2 is an open-source model series released by Google in 2024. The 2B parameter version targets resource-constrained environments and can run on ordinary laptops or small workstations. It inherits the Gemma series' dual-layer attention mechanism and knowledge distillation training method, performing stably on tasks like intent recognition, text classification, and short conversations—though complex mathematical reasoning and long-chain logic are not its strengths. CounterDeal chose it precisely for its lightweight, offline-deployable nature and its sufficiency for navigation and retrieval tasks.

Deterministic RERA Auditor

Compliance checks bypass the LLM entirely, using TypeScript to implement strict mathematical calculations that validate against MahaRERA's floor loading limits and statutory fees. This is the most important safety valve in the entire system: since regulatory numbers never pass through a language model, there is no possibility of them being "fabricated."

MahaRERA is the real estate regulatory authority established by the Indian state of Maharashtra under the 2016 Real Estate (Regulation and Development) Act. It is responsible for registering real estate development projects, broker registrations, and handling consumer complaints in Pune, Mumbai, and other locations. Its regulations contain detailed provisions on floor space index (FSI, the ratio of built-up area to plot area), statutory fees, pre-sale conditions, and more; violations may result in project suspension or heavy fines.

In real estate transactions, floor loading limits directly determine sellable area and pricing, while statutory fees affect the total transaction cost. These numbers are legally binding and must be cited precisely. CounterDeal hardcodes MahaRERA rules into TypeScript calculation logic, making every compliance check reproducible and auditable—fundamentally different from relying on a language model to "recall" regulatory clauses. For brokers, a calculation path that can be shown to clients or regulators carries more legal credibility than natural language explanations generated by a model.

TabPFN Deal Rescue Engine

Counterfactual deal probabilities and concession recommendations are computed by a local FastAPI/Uvicorn microservice built on Prior Labs' TabPFN-3.5 tabular foundation model. Tabular models typically outperform general-purpose LLMs on structured data, and they too run entirely offline.

TabPFN (Tabular Prior-Data Fitted Network) is a tabular foundation model proposed by Prior Labs. Its core idea is to pretrain a Transformer on large amounts of synthetic tabular data so that it learns to "learn in context"—given a small batch of training samples and samples to predict, the model can directly output predictions without gradient updates. TabPFN-3.5 is the latest version of the series, optimized for small-to-medium-scale tabular data (typically hundreds to thousands of rows), often matching tuned gradient-boosted trees on classification and regression tasks with extremely fast inference.

Unlike general-purpose large language models, TabPFN is designed specifically for structured numerical data and does not generate free text, so it won't "fabricate" values outside its training distribution. In CounterDeal, it is wrapped as a local FastAPI microservice that receives deal features (such as budget, location, unit type, and market heat) and outputs counterfactual deal probabilities and concession recommendations. This "specialized model for specialized predictions" approach is the concrete implementation of the "tabular model handles structured prediction" component in the neuro-symbolic architecture.

Frontend

Next.js (App Router) + React + Tailwind CSS, paired with a matte dark design system optimized for brokerage workflows. The technology choices lean conservative, but for local deployment scenarios, this is a reasonable choice.

Why Open Source and Local AI Are Irreplaceable in Such Scenarios

The author lists two core reasons, worth expanding on:

Zero data leakage. Clients' purchase requirements and financial inquiries never leave the machine, completely eliminating third-party server exposure. This isn't a matter of "better privacy policy"—architecturally, there is simply no leakage path.

Absolute legal precision. Compliance checks run through deterministic code rather than a language model, verifying statutory numbers item by item. Under regulatory frameworks like MahaRERA, an auditable code path is far more reliable than "the model says it got it right."

Real-World Validation: A 30-Day Pilot with The Real Estate Club in Pune

To validate whether the system could survive in real business operations, the author collaborated with Gaurav Khatri, operator of the local Pune company The Real Estate Club, on a 30-day pilot.

Pain Points: Lead Loss and Compliance Anxiety

Gaurav's daily life was flooded with WhatsApp messages. Real estate client messages are a chaotic mix of budgets, locations, and vague intentions, scattered across hundreds of chat records. Hot leads were lost directly because they were buried under cold ones. Meanwhile, every time a client asked about a specific property, he had to manually check MahaRERA regulations and floor loading limits—an extremely time-consuming process. And his fear of general-purpose AI was very concrete—what if the model got a single statutory number wrong?

Solution: Multiple Agents Each with a Clear Role

The pilot adopted a low-cost, fast multi-agent structure, with each bot having a single responsibility:

• Concierge: A fast parser reads incoming chats and extracts actual requirements like budget, location, and unit type.
• Qualifier: Tags leads as hot, warm, or cold based on urgency of tone, so Gaurav knows who to call first.
• RERA Auditor: Connects to a hardcoded local database rather than AI guesswork, verifying actual Pune property regulations with zero hallucination.

Results: Three Quantifiable Benefits

• 40% increase in conversion rate: Hot leads went from "discovered days later" to "surfaced within minutes," letting Gaurav lock in deals before buyers turned to competitors.
• Compliance checks cut from 45 minutes to under 3 seconds: Regulatory verification that once required manual lookups is now nearly instantaneous.
• Over 180,000 rupees in additional revenue: The system rescued high-value deals that would otherwise have been buried in chat records forever.

These numbers come from a single pilot with a single brokerage—a limited sample, but with clear directional significance: in compliance-intensive industries, moving the LLM away from areas it doesn't excel at yields immediate results.

A Pattern Worth Noticing: The Practical Value of Neuro-Symbolic Architecture

CounterDeal's significance extends beyond real estate brokerage itself. It demonstrates a reusable architectural pattern: use small local models for language, deterministic code for rules, and specialized tabular models for structured prediction, with the three collaborating through clear interfaces.

For domains like finance, law, and healthcare—where "saying one wrong sentence is costly"—this separated design is more pragmatic than pursuing a single powerful model. Open source also allows this solution to be audited, modified, and ported to other regulated industries.

The project is open-sourced on GitHub (smtherer/counter-deal), including local deployment instructions and a demo video. The author invites developers to clone the repository and test it, and welcomes suggestions for improving the tabular microservice.

Neuro-symbolic AI refers to a methodology that combines neural networks (strong at perception, language understanding, and pattern recognition) with symbolic systems (strong at logical reasoning, rule execution, and interpretable computation). Its motivation: pure neural networks are flexible but lack verifiability, while pure symbolic systems are precise but struggle with natural language and fuzzy input. Neuro-symbolic systems let each do its job and collaborate through clear interfaces.

In CounterDeal, gemma2:2b handles parsing user intent and conversational navigation (the neural part), the TypeScript RERA auditor performs statutory calculations (the symbolic part), and TabPFN handles structured probabilistic prediction (arguably a learned symbolic model). This separation brings three benefits: the language model never touches regulatory numbers, eliminating hallucination risk; calculation paths are auditable and reproducible; and each component can be independently replaced or upgraded. For regulated domains like finance, law, and healthcare, neuro-symbolic architecture offers a more pragmatic, more controllable path to deployment than "pursuing bigger models."

Frequently Asked Questions

Does CounterDeal need an internet connection? No. AI inference, compliance calculations, and the deal probability engine all run locally.

Why not use a large model like GPT-4 for compliance? Because regulatory numbers don't allow hallucinations. Deterministic code can be audited and reproduced; language models cannot.

Is gemma2:2b sufficient? It's sufficient for tasks like intent recognition and conversational navigation; complex reasoning is deliberately excluded from the model's responsibilities.

Can this solution be ported to other cities or industries? The architecture itself is general—just replace the regulatory database and business rules. It applies to other regulated local service industries.

One-line Summary

CounterDeal uses neuro-symbolic architecture to separate language models from deterministic code, delivering a local real estate compliance workbench with zero leakage and zero hallucination.

Paragraph Summary

CounterDeal is a local neuro-symbolic workbench built by developer smtherer for real estate brokers in Pune, India, addressing two major pain points: sensitive high-net-worth client data and regulatory calculation hallucinations. The system splits tasks: gemma2:2b runs locally via Ollama for intent recognition and conversational navigation; a deterministic RERA auditor implemented in TypeScript hardcodes MahaRERA regulations to perform compliance checks without a language model; a TabPFN-3.5 tabular model microservice computes counterfactual deal probabilities and concession recommendations; and the frontend uses Next.js and Tailwind CSS. A 30-day pilot showed hot lead identification shortened from days to minutes, conversion up 40%, compliance checks cut from 45 minutes to under 3 seconds, and over 180,000 rupees in rescued revenue. The author believes the architecture can be ported to regulated industries like finance, law, and healthcare, with the core idea being to keep LLMs away from precise calculations.