How we built Dealzzz — a vehicle auction platform that runs entirely on WhatsApp, powered by contextual AI agents with memory, guardrails, and human fallback.
India’s used two-wheeler market runs on relationships. Dealers talk to each other, share inventory photos, negotiate prices — and 99% of this happens on WhatsApp. When we looked at how to build an auction platform at DriveX, the conventional approach would have been a dedicated app or web portal. But we kept running into the same friction: dealers don’t want another app. They already have a workflow, and it lives in WhatsApp.
So we asked a different question: what if the auction platform was WhatsApp?
Not a bot that sends links to a website. Not a notification system that redirects to an app. An actual auction — browsing, bidding, winning, paying — conducted entirely through chat.
The result is Dealzzz, India’s first auction platform built natively on WhatsApp, powered by agentic AI.
The numbers made this decision obvious:
The hard part isn’t choosing WhatsApp. It’s building a real-time, multi-user auction experience within the constraints of a chat interface.
Dealzzz isn’t a rule-based chatbot with decision trees. It’s built on AI agents that understand context, maintain memory, and make decisions autonomously — while keeping humans in the loop for edge cases.
Our primary focus from day one was interpretability, transparency, and human fallback. In an auction setting, where real money changes hands, the AI’s actions need to be visible and comprehensible. You can’t have a black box deciding who wins a bid.
This is where things get interesting. A stateless chatbot can answer questions. An agent with memory can run an auction.
We built a two-tier memory system:
Short-term memory tracks the immediate auction session — what vehicles the user has viewed, bids they’ve placed, questions they’ve asked in the last few minutes.
Long-term memory persists across sessions — bidding history, vehicle preferences (this dealer prefers scooters over motorcycles), location, typical spending capacity, preferred languages. When a returning user starts a new auction session, the agent already knows their context.
This is what makes the experience feel conversational rather than transactional. The agent doesn’t ask you the same questions twice.
We run multiple LLM providers — OpenAI, Claude, Gemini — behind a unified gateway that handles:
Prompts are code. We treat them that way.
We built a prompt management system with versioning, A/B testing, and rapid deployment. When we need to change how the agent handles a specific auction scenario — say, how it responds to a bid that’s below the reserve price — we can ship a prompt change, test it against a control group, and measure the impact on auction completion rates.
AI agents in production need boundaries. Ours include:
The agents handle most conversations autonomously. But in edge cases — disputes, unusual requests, situations the agent isn’t confident about — the system seamlessly hands off to a human operator. The operator sees the full conversation history and context, so the user doesn’t have to repeat themselves.
The handoff is invisible to the user. They’re just chatting. Sometimes the responses come from an agent, sometimes from a human. The experience is the same.
A traditional auction platform has a grid of vehicles, countdown timers, bid buttons, price charts. We had a chat window. No custom UI components, no rich interactive elements — just text, images, and WhatsApp’s native message types.
We had to reimagine every auction interaction as a conversation:
This was a non-obvious challenge that would break most off-the-shelf NLP:
We built custom parsing and understanding layers specifically for Indian auction contexts. The agent understands regional number conventions, mixed-language queries (Hinglish is the norm, not the exception), and the implicit context that comes with India’s dealer communication style.
India’s dealer network spans Tier 1, 2, and 3 cities. A dealer in Chennai communicates differently from one in Jaipur. We built in language detection and regional support to make the experience feel local — not like a translated English product, but like a platform that was built for that language from the start.
Every decision the agent makes is logged: what it understood, what options it considered, what it chose, and why. This gives us two things:
Building Dealzzz taught us that the hard part of production AI isn’t the model — it’s everything around it. Memory management, prompt versioning, multi-provider routing, guardrails, human fallback, cultural context, logging. The LLM is maybe 20% of the system. The other 80% is engineering.
The future for AI-driven commerce in India is wide open. WhatsApp is where the users are. AI is mature enough to handle complex workflows. The question isn’t whether conversational commerce will happen — it’s who builds the infrastructure to make it reliable, transparent, and human-centric at scale.
Dealzzz is our answer.