Converge

Framing the Problem

Conferences concentrate a huge amount of value — interesting people, novel projects, and hard-won knowledge — into a few days. Yet attendees spend most of their energy on logistics instead of on the things that actually matter: which session to attend, how to get there, how to remember what they saw, who they should meet, and how to stay in touch afterwards.

The result is wasted potential. People photograph slides they never revisit, most audience questions go unanswered, networking happens by luck rather than intent, and the conversations and connections evaporate the moment the event ends. This affects everyone at the event: attendees who leave without the connections or knowledge they came for, speakers whose ideas stop spreading once they leave the stage, organizers who struggle to build lasting community, and sponsors who can only broadcast rather than participate.

Existing conference apps solve scheduling. They do not solve learning, networking, or community.

Idea Explanation

Converge is a people-first conference companion. Instead of treating people as metadata attached to talks, it makes people and the projects they're building the primary objects of the platform, and it focuses on the moments where value is actually created or lost.

It fixes the problem through a few core ideas:

• People before sessions & projects before companies — rich profiles centered on what someone is building and what conversations they want (and don't want), so networking becomes intentional instead of random.

• Moments instead of photos — a one-tap Session Moment captures the timestamp, transcript, current slide, and context, replacing the dead-end of photographing slides.

• Questions become conversations — Q&A turns into persistent discussion threads that can continue from question → answer → follow-up → meetup, long after the room empties.

• Conferences gain memory — a living knowledge graph links people, sessions, projects, questions, and discussions so the value of the event remains searchable after it ends.

• An AI Concierge that answers "who should I meet?", "what did I miss?", and "find projects related to my interests" on top of that graph.

The throughline: the conference ends, but the network remains.

Implementation

The architecture is a single full-stack TypeScript app running on Bun, built with TanStack Start for server-side rendering and routing, with TanStack DB providing a reactive client-side data layer that syncs against the backend. The frontend communicates with backend server functions/endpoints for reads and mutations, and live experiences (session timelines, live Q&A, discussion updates, billboard) are pushed to clients over Server-Sent Events rather than polling.

• Database: A PostgreSQL database is the system of record, modeling the core objects — people, sessions, projects, moments, questions/discussions, and the relationships between them that form the knowledge graph. TanStack DB sits in front of it on the client for optimistic, reactive UI.

• Auth: Better Auth handles identity, with passkeys plus GitHub and LinkedIn social login — a natural fit for a developer/professional conference audience.

• AI & agents: Every major object is also exposed through an MCP server with OAuth integration and MCP-Apps support from day one, so the AI Concierge and external AI agents can act on a user's behalf (read sessions, find people, manage notes/bookmarks/contacts).

• UI: Built from the free Progress KendoReact components, with interaction polish from the make-interfaces-feel-better and transitions.dev skills

• Hosting: Deployed on Railway using config-as-code, with the Postgres database hosted alongside the app.

Challenges

• styling the Kendo textarea component

• learning 6 different new technologies

Accomplishments

• Scope shipped

  • Full-stack conference companion: 19 domain tables, 36 routes (16 app pages + 14 REST endpoints + 2 OAuth discovery + MCP), ~16k lines of TypeScript/TSX across 126 files — production-deployed at converge.sams.land.

  • Every core feature working end-to-end: AI search, moment capture, shared-moment matching, persistent Q&A → discussions → meetups, people discovery with intent filters, connections, messaging.

  Technical depth (the stuff judges reward)

  • MCP platform with OAuth 2.1 — ~29 tools across people/projects/personal/conference, including interactive MCP Apps (_app variants returning UI resources). Every object in the app is also an agent-callable API. The .well-known OAuth discovery routes are implemented, not stubbed.

  • Real-time optimistic UI — TanStack DB client collections + SSE event bus (/api/stream), so writes (votes, moments, connections) reflect instantly and sync across clients.

  • Passkey + GitHub auth via better-auth, with server-side session enforcement on every server fn and API route.

  • Natural-language AI search over live conference data (people, sessions, projects, moments) — pluggable provider, runs on claude-sonnet-4-6.

  • Solved a genuinely hard build problem: keeping the server-only pg driver out of the client bundle through TanStack Start's server-fn dependency analysis (documented in commits — shows engineering rigor under time pressure).

  Originality / product

  • Shared-moment matching — deriving warm introductions from people who bookmarked the same timestamp in a talk. This is the novel idea; nobody else turns attention-proximity into a connection signal.

  • A coherent thesis carried through the whole data model: people before sessions, projects before companies — sessions/Q&A/moments all hang off people and projects as the primary objects.

  Impact framing

  • Attacks all three classic conference failure modes at once: discovery (find the right people), attention (stop photographing slides), continuity (questions don't die when the speaker leaves).Next Steps

The long-term goal is to turn conferences from temporary events into persistent, searchable communities. Concrete next steps beyond the MVP:

• Deepen capture & recall — voice notes, AI-generated session highlights and summaries, and richer transcript/slide linking within Session Moments.

• Make networking smarter — intent- and availability-based people matching, AI-scheduled coffee chats, and a Project Explorer with trending/categories/creators.

• Grow the community layer — topic channels, meetup suggestions, speaker responses, and the in-venue Conference Billboard that surfaces people and ideas instead of sponsor logos.

• Expand the MCP platform — broaden the Conference, People, Project, and Personal MCP surfaces so third-party agents can do more on attendees' behalf.

• Persist beyond the event — keep the knowledge graph and connections alive after the conference, so the network and its knowledge stay useful between events.

• build ui for big screens