The problem DealForge solves

Procurement is broken for small and mid-sized businesses. Getting competitive supplier quotes means hours of phone tag, inconsistent negotiation, and institutional knowledge that lives only in someone's head or gets lost entirely when that person leaves.
The status quo looks like this: a buyer manually calls three to five suppliers for every order. Each negotiation is only as good as the person making the call their mood, their memory of past deals, whether they remember to mention a competing quote. Supplier pricing history, negotiation tactics, and leverage points are stored in spreadsheets or nowhere at all. There's no verifiable record of what was agreed and why. The whole process can take days and still produce a mediocre deal.

Procurement Negotiator iNFT replaces that process with an AI agent that makes real phone calls and negotiates for you.
You describe what you need and the agent dials the supplier, introduces itself, anchors on a low opening price, drops competitor quotes as leverage, and pushes for a final commitment. All via a live voice call, powered by 0G Compute LLM inference.

More importantly, the agent gets smarter with every call it makes. After each negotiation it writes what it learned (the supplier's best price, their willingness to negotiate) into its brain, a document stored on 0G Storage with the hash committed on-chain. The next time it calls that supplier, it already knows their floor price and uses competing quotes from other suppliers it has previously called as leverage. The agent compounds its knowledge the same way a seasoned procurement professional does, except it never forgets and never leaves the company.

Because the agent is an iNFT on 0G Chain, that accumulated intelligence is a verifiable on-chain asset. The agent is owned by whoever holds the NFT. You can mint one, train it across hundreds of negotiations until it has deep supplier relationships and hard-won pricing intel baked into its brain, and then sell access to other businesses — or sell the NFT itself, transferring all of that accumulated knowledge to the new owner. Usage is sold through an on-chain credit system: buyers purchase credits to run RFQs through your agent, and the fee flows directly to the NFT holder's wallet. The smarter your agent gets, the more valuable it becomes, and the more you can charge for access to it.

Challenges we ran into

The biggest technical hurdle was making the voice negotiation actually work with the model available on the 0G Compute testnet. We were constrained to Qwen 2.5 7B Instruct a small model that doesn't have the reasoning depth you'd want for a multi-turn negotiation that needs to track what step it's on, remember what the supplier just said, and respond naturally in one or two sentences without going off-script.
The first version of the agent was a mess. It would skip steps, repeat itself, ignore what the supplier said entirely, or produce responses so long that the conversation felt robotic. Getting a 7B model to reliably follow a multi-step negotiation script, anchoring on price, dropping a competitor quote at the right moment, pushing for a final offer and then closing, required a lot of prompt engineering iteration.

We eventually landed on a very rigid step-by-step system prompt that told the model exactly what to say at each stage, almost word for word. Less creative, but far more reliable with a small model. The challenge was finding the right level of specificity, too loose and the model wandered off script, too rigid and it stopped responding naturally to what the supplier actually said.

Use of AI tools and agents

The core of the project is a voice AI agent that negotiates procurement deals over real phone calls. When a buyer submits an RFQ, the agent initiates an outbound call via Twilio to the supplier's phone number. Twilio's ConversationRelay connects that live call to a WebSocket server where the agent processes each turn of the conversation in real time.

Every time the supplier speaks, the transcribed text is sent to 0G Compute, 0G's decentralized inference network, where a Qwen 2.5 7B model generate the agent's next response. That response is streamed back token by token to Twilio's text-to-speech engine, which reads it aloud on the call. The result is a fully autonomous back-and-forth voice negotiation with no human in the loop.

After the call ends, the agent uses 0G Compute again to extract the structured quote details, unit price, minimum order quantity, and lead time, from the negotiation transcript. That transcript and quote are then uploaded to 0G Storage and committed on-chain, making every negotiated price fully verifiable.