Inspiration
Every VC associate we talked to has the same Monday morning. Their LinkedIn and email inboxes are full of founder pitches, most of them generic. They check each sender by hand, then write five to ten one-page briefs a week for their partner. And the good founders who show up at the wrong time, for example while the fund is raising from LPs, get parked and forgotten.
We interviewed 20+ people, including a VC associate at a consumer and CPG fund. She told us her LinkedIn inbox "is just too much," described her exact weekly brief, said she would trust an AI filter, and gave us a real "park until we raise" example. Her advice was to "combine tracking with filtering." That became Last Touch.
What it does
Last Touch is an agent layer that sits next to the tools a fund already uses. It does not replace the CRM.
- Associate agent: reads inbound founder messages, extracts the company, stage, traction, team, ask and mutual connections, checks the company in Akta, and scores the founder with fixed rules. Good founders get a one-page brief in the partner's format. The rest get a logged reason: generic, off-thesis, or not VC-scale.
- Partner agent: applies the partner's written criteria and returns meet, pass or park, with a one-line reason.
- Watch agent: monitors parked founders using Akta news and signals plus a fund-status flag. When a founder raises, hires, makes news, or the fund closes, it adds "Follow up now: " to their note and shows an alert.
Every claim in a brief links back to its source, either the original message or an Akta field.
How we built it
- Nicrron API for every model call: extracting fields from free text, classifying sector and stage, and writing briefs and short reasons from extracted fields and Akta data only.
- Roundtable Notes as the shared state between agents, with one note per founder.
- Akta API for company lookup, enrichment and news signals.
- PRISM (prismtrace-sdk) to trace every agent run, so each output can be traced back step by step.
- Plain code, not AI, for the decisions that matter: thesis match on sector and stage, a mutual-connection boost, a generic-message detector, and a weighted final score with a visible breakdown.
- Stack: Vercel, Roundtable, GitHub, Nicrron
The model never invents numbers. It only reads and writes from fields we extracted or fetched.
Challenges we ran into
- Scope: we first planned to read WhatsApp and iMessage. A VC we spoke to at the event said their firm keeps everything on email and Slack and would be uncomfortable with an agent reading personal chats, so we dropped them and focused on the inbox.
- Trust: a filter that silently drops a good founder is worse than no filter. We kept the thesis checks as fixed rules, showed the score breakdown, and logged a reason for every drop.
- Demo data: we used 14 synthetic messages with fictional people. Where we needed a real Akta lookup, we used well-known public companies.
Accomplishments that we're proud of
- A working end-to-end run on 14 inbound messages: 5 dropped with a reason each, the rest ranked by each person’s own criteria.
- Agents resolve who owns each relationship: Ben’s agent tells Ava’s agent he met Maya 46 days ago with no follow-up, so Ben re-intros and Cara takes the first call.
- Briefs follow the format a working associate uses, and every claim is clickable back to its source.
- The watch agent turns a parked founder into a "follow up now" alert when something changes.
What we learned
- The pain we heard was specific and behavioral: weekly briefs, hand-checked senders, parked founders.
- Funds do not want another CRM. The product has to be a layer.
- Trust in an AI filter comes from visible rules and sources, not from a better model.
What's next for Last Touch
- A small MCP server, so a partner can ask Claude "which parked founders just raised?"
- Sync to the CRM a fund already uses, such as Folk or Attio.
- Referral routing: offer an off-thesis founder to another fund's agent through a shared note.
- Pilot with the funds we interviewed today, then test pricing per seat against what funds pay for CRM seats.
- Connect real inboxes with fund approval, after the demo on synthetic data.## Inspiration
Log in or sign up for Devpost to join the conversation.