WARMTH: A Social Graph Engine for Live Conversations

Warmth started with a frustration from Web Summit last year. I'd done the intentional thing: scraped the attendee directory, pulled profiles, built a connection strategy in a spreadsheet, sent careful follow-ups. It worked, but all the effort sat around the conversation, not in it. I was reconstructing meetings afterward instead of being present during them. The question that became Warmth: what if the structure could come from the conversation itself?

The thesis is that the most valuable connections in go-to-market aren't only the most monetarily profitable ones. They're the ones that are professionally and personally profitable too, and those are exactly the ones a CRM field can't capture. Warmth is an attempt to build a smarter social network for your GTM strategy, one that records how you connect with someone, not just that you did.

Every technical decision came back to that intent, starting with privacy. The microphone is on your phone, so the architecture had to earn that trust. Warmth listens passively, does all speech recognition and entity extraction on-device, and never sends raw audio anywhere. What leaves the phone is an encrypted JSON payload of structured signal, sent to your own backend, not a stream of your conversations to a third party. The two-tier split (phone proposes, backend decides) is as much a privacy boundary as an engineering one.

From that signal, Warmth builds a per-person model that evolves in real time. Each 30-second window produces a context delta (topic weights, communication style, values, pain intensity) that folds into a PersonNode, so the knowledge graph is constantly updated as the conversation unfolds rather than reconstructed at the end. By the time you walk away, you have a narrative, not a bag of keywords: what this person cared about, what they kept returning to, where their real pain was.

The two scores are the part I find most interesting. Before the conference, the model predicts how the relationship might go. After the meet, it produces an actual warmth score. The gap between them does two jobs. First, it's a check on the model: how well did our pre-meet read of this person hold up against the real conversation, and where should the prediction improve? Second, and more usefully, the post-meet warmth tells you something human, how well you actually got on, whether this is someone worth investing in, or whether the better move is a warm intro to someone in your founder community who'd be a stronger fit. That routing decision (reconnect yourself, or hand off) is the core of the loop.

Underneath, capture runs on-device (NLTagger NER, relationship regex, ICP keyword proximity) with a Soniqo wake-word path and 30-second capture windows. Warmth deliberately doesn't own ICP fit (Zero does); it builds relational warmth as a separate dimension on top, because a perfect ICP match can still be a cold conversation and an off-ICP person can become your best introduction. I explored TGN, GCN, K-Means, and HDBSCAN for the graph layer and shipped a NetworkX knowledge graph connecting each person to their interests, topics, values, and pain points, rendered as a radial graph on the dashboard.

This was also a chance to design with hardware in the loop again, Apple Watch and iOS as a capture surface rather than just a screen. The bigger bet behind all of it: AI should make us better at connecting, not replace the connection. Warmth captures the human moment, structures it without flattening it, and hands it back to you.

Built in a day at the GTM Hackathon, London, June 2026.