Shoppers search, discover, and buy by talking.

Rezolve AI lets shoppers search, discover and buy by talking. Terrafusion built the multi-agent harness underneath: a planner and reflect loops on LangGraph, product search as a tool over a graph knowledge base, three memory stores, duplex and cascaded voice, and an eval loop that feeds post-training of Rezolve’s models. First production in twelve weeks; still shipping two years on.

Client Rezolve AI
Sector SaaS — retail & e-commerce (Nasdaq-listed)
Engagement Conversational commerce platform — 2+ years, ongoing
Approach Multi-agent harness — planner, memory, voice, eval

The approach

Scroll diagram horizontally →

langgraph, the runtime shopper voice, image, text planner plan, replan observe reflect calls rezolve llm post-trained on rezolve data graph search tool, product graph memory short-term: this turn long-term: this session episodic: this shopper duplex voice interrupt, it keeps up cascaded voice pipeline, tuned tts eval scores every action post-train weeks 0 2 4 6 8 10 12 wk 12 first production
Illustrative. One request through the harness: planner, graph-search tool, three memory stores, two voice paths that re-enter eval, and a post-training loop back into Rezolve’s own models. The ruler is weeks to first production.
12 Weeks to first production
5+ Products in production on the platform
2+ Years embedded, still shipping

Rezolve’s models, post-trained on Rezolve’s data. The planner, memory stores, voice layer and eval loop are the Terrafusion-owned core, on clusters we deploy and operate in Rezolve’s cloud. Voice models fine-tuned for commerce on Microsoft infrastructure, with co-authored research papers on voice.

Key outcomes

Platform

One platform, many merchants

Products run in production for merchants from Harrods to M&S on the same platform, configured per merchant rather than rebuilt.

Voice

Shoppers can interrupt mid-sentence

Two voice paths, duplex and cascaded, each a full harness under the same eval loop, so a shopper can talk over the system and it keeps up.

Still shipping

From proofs of concept to live merchants

Two years embedded: proofs of concept became signed clients and live deployments for a Nasdaq-listed company.

The challenge

Rezolve AI set out to let shoppers search, discover and buy through natural conversation — voice, image and text. That takes product search across real inventory, agents that remember context across sessions, and voice that can carry a commerce conversation.

Off-the-shelf tools fell short: agents forgot the session, voice as a wrapper around hosted text-to-speech was rigid, and every merchant deployment risked becoming a bespoke build. One platform had to serve them all.

Harnesses and evaluation: the Software Factory

Harrods and M&S are Rezolve AI’s merchants, not Terrafusion’s clients.

Start with a 30-minute scoping call.

Start a conversation