Team
A2Atrium
Project Concept
An A2UI-native surface for the moment of meeting — when you encounter someone, the agent generates the interface you need around what you’re trying to do with them, not around what a CRM thinks a contact page should look like.
Track:
Kill the Dashboard (with a remix of The Copilot That Ships)
Entry
Status: Submitted
Last saved: May 09 at 6:00 PM PDT
Team Roster
Message board not available for this team yet.
Michael Lopez Team Lead RSVP Approved
Founder at Inner Cartography
**Solo build by Michael (InnerCartography).**
**CopilotKit / AG-UI:** Wired the CopilotKit provider and `useFrontendTool({render})` registration for the controlled-tier ContactCard component. Used AG-UI's durable conversation thread and agent ↔ frontend state sync to keep the canvas and agent in sync across intent selections.
**A2UI declarative layer:** Designed the component catalog (DraftMessage, TimeSlots, ResonanceMap, CollaborationProposal, RememberAction, ContextBridge, PortalLink, and others) and the JSON spec contract the agent emits. Built the spec renderer that resolves agent output against the catalog — the guardrail that prevents the agent from inventing arbitrary UI.
**LangChain Deep Agents + Gemini:** Integrated LangChain's deep agent orchestration as the planning and dispatch layer. Used Gemini 2.5 Flash as the LLM (via the `google-genai` SDK) for both vision (LinkedIn screenshot extraction) and generation (follow-up drafting with explicit gold-extraction step, convene micro-KG generation, remember annotation).
**Telegram conversational layer:** Built the full Telegram bot from scratch — photo ingestion with Gemini Vision, inline keyboard contact selection, `/followup` with two-step prompting (gold extraction → personalized draft → Gmail compose URL), `/convene` for person×person and person×topic, `/remember` with pending state machine. Bot writes specs back to `sample-runs/` so the canvas reflects the same curated content.
**Frontend (Next.js 15):** Built the contact sidebar, intent button row, main pane spec renderer, place-aware theming system, 15-second polling for live contact updates, and the vesica portal modal on the Convene surface.
**All infrastructure, architecture, and product design were solo.**
Michael builds at the intersection of spatial computing, knowledge graphs, and agentic UX under the InnerCartography brand. His project ROOM won first place (PLACES track) at the RP1 Open Metaverse Hackathon. He’s currently building Hermes, an AI agent and knowledge graph for managing high-signal IRL relationships through tiered generative UI on Telegram. Stack: Supabase, DigitalOcean, Netlify. He publishes hyperblogs at innercartography.one and learned to code backwards — through projects, not curriculum.
Knowledge Graphs & Agentic Systems — ~1 year, shipping
Building Hermes (relationship/event graph on Supabase + Telegram) and PKOS (applying ROOM to property management). Learned backwards — through projects, not curriculum.
Generative UI — Active focus
Tiered, graduated-access interfaces grounded in real relational tiers from property work.
Room - spatial memory and intelligence for buildings
PathfinderXR- a tool for caseworkers to offload cognitive complexity
Chorus - intergenerational care platform
PKOS - Personal Knowledge Operating System
Hermes — an AI agent + knowledge graph for managing high-signal IRL relationships and event contexts. Core stack: Supabase, DigitalOcean, Netlify, Telegram as the primary IO surface. The architecture is “ephemeral surfaces with persistent edges”: event-themed landing pages that dissolve