Introduction & The Problem
When you are handling high-net-worth client financial records and strict property regulations, standard cloud-based AI tools introduce massive data privacy risks. On top of that, language models are notoriously bad at hard math—which is a disaster when you are dealing with strict statutory compliance like MahaRERA regulations.
I wanted to solve this real-world problem for real estate brokerages in Pune, so I built CounterDeal: a local-first workspace that completely removes cloud data leakage and LLM hallucinations.
📺 Video Demo
Here is a quick walkthrough of the terminal in action:
How I Built It (The Tech Stack)
CounterDeal uses a hybrid, neuro-symbolic architecture to split the workload between natural language and hard deterministic logic:
- Local AI Runtime: Running gemma2:2b locally via Ollama to handle natural language workspace navigation and the Buyer Concierge safely with zero data leakage.
- Deterministic RERA Auditor: Bypasses LLMs entirely for legal and compliance checks, running strict TypeScript math against MahaRERA loading limits and statutory costs to prevent any hallucinations.
- TabPFN Deal Rescue Engine: A local FastAPI/Uvicorn microservice running Prior Labs' TabPFN-3.5 to calculate counterfactual deal probabilities and concession recommendations.
- Frontend: Next.js (App Router), React, Tailwind CSS, and a custom matte dark design system optimized for professional broker workflows.
Why Open-Source & Local AI Matter Here
By keeping the entire stack local and open-weight, CounterDeal achieves what closed commercial APIs can't:
- Zero Data Leakage: High-net-worth client buyer requirements and financial queries never leave the machine. Everything runs locally, eliminating third-party server exposure.
- Absolute Legal Precision: Compliance checks bypass the language model entirely and run through strict, deterministic code to verify statutory figures.
🚀 Real-World Validation: The Real Estate Club (Pune)
To see if this system could actually survive the real world, I teamed up with Gaurav Khatri, who runs a local firm called The Real Estate Club here in Pune.
The Mess: Lost Leads & Compliance Paranoia
Gaurav was basically drowning in WhatsApp messages. If you've ever dealt with real estate clients, you know their texts are a chaotic mess of budgets, locations, and mixed signals scattered across miles of chat history. He was straight-up losing hot leads simply because they got buried under cold ones.
On top of that, every time a client asked a specific property question, Gaurav had to manually cross-check strict MahaRERA regulations and building loading limits. It was a massive bottleneck—and he was terrified that a generic AI would hallucinate a statutory number and land him in legal hot water.
The Fix: Getting the Bots to Work Together
We set up a quick multi-agent session using an ultra-cheap, fast model architecture, giving each bot a highly specific job:
- The Concierge: A fast parser bot read incoming chats to instantly extract what the client actually wanted (budget, location, BHK).
- The Qualifier: A second bot looked at the urgency of the tone and immediately flagged them as hot, warm, or cold so Gaurav knew who to call first.
- The RERA Auditor: Instead of letting an AI guess compliance numbers, we hooked this bot up to a strict, hard-coded local database. It checked Pune's actual property regulations with zero hallucinations.
The Payoff
We ran this as a 30-day pilot, and the results blew both of our minds:
- 📈 40% Spike in Conversions: By pulling hot leads out of the WhatsApp graveyard in minutes instead of days, Gaurav was locking down serious buyers before they went to competitors.
- ⏱️ Seconds, Not Minutes: Compliance checks that used to take 45 minutes of manual digging dropped to under 3 seconds.
- 💸 ₹1,80,000+ in Extra Revenue: The system rescued high-value deals that literally would have been lost forever in messy chat threads.
Run It Yourself & Let’s Connect!
If you want to test this locally, review the code, or run it on your own machine, everything is open-source:
GitHub Repository: https://github.com/smtherer/counter-deal
YouTube Demo: https://youtu.be/ECRpxFGKpdw
Want to test or collaborate? Clone the repo, check out the local setup instructions in the README, and let me know if you run into any issues or have ideas for improving the tabular microservice. Drop a comment below!
Top comments (0)