Skip to main content
NurraJoin Nurra’s waitlist

The technology

An assistant that helps you get things done.

Nurra connects AI understanding to the work of a local marketplace. A conversation can become a search, a listing draft, or a proposed arrangement. Each action has its own tools and checks.

Your words + your contextA useful next step

Follow one action

Watch a message become a useful draft.

Choose a step. See what the model, agent, app, and person each contribute.

You → your assistant
Your product photo

Help me sell this used chair for $60.

The model makes sense of what you mean.

The model reads your message and photo together. It understands the goal and the details you have supplied.

Missing details are questions to ask, not facts to invent.

This is the architecture of one listing action. Searching, editing, and arranging an exchange use their own tools and checks.

The model layer

Clear roles. Room to evolve.

Each part has a job. The model brings understanding; the agent connects it to the product.

Broad understanding, vision and writing

GPT-6 Luna

The current source default for Home, Discovery, and listing work. It interprets requests and photos, handles corrections, and writes replies and listing copy.

Selected, bounded decisions

JEV

Used for specific finite decisions, including selected discovery-relevance and listing-moderation checks. It does not replace the assistant’s general understanding.

Development architecture, September 2026. These are source defaults and assigned roles; model configuration can change as we test.

The cost of running Nurra

One turn. One listing.

Two ways to measure our AI costs, based on what the assistant helps you do.

Home assistant

Average cost per turn

Awaiting test results

Everyday help and finding what you need, measured together.

Listing creation

Average cost per completed listing

Awaiting test results

The full creation flow, including follow-up messages and corrections.

How we’ll measure this

For Home, we’ll divide the recorded AI cost across tested Home turns by the number of turns. A turn is one message and the assistant’s response or action. Finding products and services is part of the same assistant.

For listings, we’ll measure the full journey from the first message or photo to a successfully created listing. That includes follow-ups, corrections, retries, and AI checks needed to finish.

Results will include the test date, sample size, and what was counted. These are our AI operating costs, not a price charged to users. Hosting and other services are separate.

The work behind the assistant

Built, tested, and refined.

Shaped through months of building, testing, and refining. We work on how the assistants understand requests, use tools, and handle corrections.

  1. Build
  2. Test everyday tasks
  3. Check the outcome
  4. Refine and repeat

We check what the assistant says and what the application actually saves.

Understanding without a scripted conversation

People should be able to explain what they mean without learning special prompts. We’re developing language support around how communities actually communicate, including messages that mix languages. The model interprets the request; the app checks details, permissions, and confirmations before acting.

Specialized agents, shared product context

Home, Discovery, and listing agents serve different parts of the experience. Each works with relevant context and the tools it is allowed to use. The aim is continuity for the person, even when the underlying task changes.

Corrections, retries, and real outcomes

“Actually, make it $90” is part of normal use. Agent work has to account for revised details, canceled work, failed requests, and repeated actions. A confident sentence is not proof that a change was saved.

Speed measured by useful work

A fast first word is not enough. We care about the time until someone has a useful search result, draft, or completed action, while preserving the facts and checks the task needs.

Testing beyond a convincing reply

We review conversation continuity, the quality of the result, and whether the application actually recorded the intended action. We are still preparing for user testing; we are not presenting a public performance benchmark.

Help us solve the next set of problems

Preparing for testing

Help shape what comes next.

Nurra is preparing for testing. If you want to build, support, or invest in this work, start a conversation with Ashfaq.

Email Ashfaq