Finding with Nemo

An agent that maps what is nearby, researches each business live with Tavily, cites every claim, and plans your route. Nemotron on Nebius, for sales, job search and sourcing and much more.

Comment

Inspiration

Deciding where to go matters more than just finding places. When I went out to find customers, there was no product that told me what was around me: which businesses are nearby, what they do, and who is worth a visit. Sales tools assume you already know who to call. Real prospecting starts earlier, in the field, and the AI inside most tools states things with no way to check them. I wanted instant research and an agent that acts, on the street and in your own space, where every claim shows its source.

What it does

Finding with Nemo is an agentic prospecting CRM built on a map.

  • Tell it what you want. A business says what it sells; an individual says what they are looking for and where. Tavily reads the public web and Nemotron drafts a profile, tags and search areas, plus standing instructions for the agent ("playbooks") shaped by what you want research to look for. You approve them, and the map opens on your city.
  • Pick an area and ask. The chat agent sees your screen (the area, the results, the client you selected), searches Google Maps for your own tags inside that area, and proposes what to save. Nothing changes until you press Approve, and new tags are remembered for next time.
  • Instant, cited research. For any client, Tavily brings back the live web and Nemotron writes a short brief. Each sentence carries a quote from its source, and my code keeps it only if that exact quote appears on the page. You only see what the page really says. The agent can do the same from chat ("any recent news on this contractor?") and tells you when it used Tavily.
  • Memory that works for your agent. Say "remember that this client wants quotes in dirhams" and it is stored in the organisation's memory, recalled by every later chat, and shown as a glowing dot in a 3D memory view that lights up when the agent uses it.
  • In the field. Plan a route that starts where you are, and record a voice note in your own language: it is translated, stored encrypted, and its facts are quote-checked.
  • Keeps learning your work. The playbooks are refreshed from what you search, save and ask.

Use cases: job seekers, sales teams and procurement. Any trade and any city, with no code changes.

How I built it

  • NVIDIA Nemotron on Nebius Token Factory, each model with a job. Nemotron 3 Ultra does careful reasoning: onboarding, extracting facts with their quotes, fact-checking. Nemotron 3.5 Super-VL runs the tool-calling agent loop, writes research briefs, drafts the playbooks and translates voice notes. Nemotron 3.5 Lightning does fast checks: tag ideas, picking route stops, and a verifier that marks any reply sentence with no supporting source as "not verified". Qwen3-Embedding-8B (also on Token Factory) powers memory search.

  • Tavily is the route to the live web: onboarding research, the cited client brief, the agent's news search, fact-checking and project finding.

  • Claude Coding for coding.

Challenges I ran into

  1. Making "cited" mean something. The Evidence Gate marks a fact verified only if its quote is found word for word in the source. In practice many model-written findings fail that check and are dropped, which is the point.
  2. An agent that does not ask what is on screen. Early versions asked the user to draw an area that was already selected, or lost track of which client was open. Sending the agent the screen state fixed it.
  3. Honest limits. Token Factory has no speech-to-text, so voice notes use the browser's recognition with a typed fallback. Token counts from a streaming bug were inflated in one evaluation run; I fixed the metering and no longer quote them. Multi-step answers can take 20 to 80 seconds.

Accomplishments I am proud of

  • One codebase for any trade and any city, with every write behind an approval card.
  • 13 of 13 prompt-injection attacks flagged, and a 30-question agent regression across three unrelated businesses that passed 30 of 30 (a regression score on the earlier models, since those questions also helped me find bugs).
  • Cited research where unsupported claims are dropped rather than shown.

What is next

Scheduled watchers as Nebius Jobs, real speech-to-text with an NVIDIA speech model, a Nemotron safety model when it is available on Token Factory, and team assignment of leads.

Built With

Share this project:

Updates

Submission history