G2 Meets

Inspiration

Technical conversations move fast. Someone mentions a colleague you have not met, an internal acronym, or a project you have never heard of. Looking it up on your phone can mean missing the next sentence.

I wanted a way to get that missing context while staying in the conversation. Instead of putting a chatbot in front of your eyes, G2 Meets shows one useful piece of context at a time. The first use case is a new engineer learning the people, projects, and language of a company.

What it does

G2 Meets combines live captions, a company glossary, and a personal relationship memory.

Finalized sentences appear on the right side of the glasses. People, organizations, and technical terms become selectable mentions. Swipe between them, then press to lock one context card on the left while the conversation continues. Press again to let the card follow your selection.

In our fictional Northstar Labs demo, a new engineer hears about Elena Rossi and ACP. G2 Meets shows Elena's role and the company's meaning of ACP, rather than guessing from the acronym alone. The desktop console shows the supporting source, confidence, and why that context appeared.

After a meeting, the app creates a title and recap. An optional private context note stays separate from the meeting transcript. Suggested relationship updates and follow-ups include their evidence and need your confirmation or rejection. Saved meetings and people can be reviewed later on the computer, not just on the glasses.

How I built it

The glasses app uses the official Even Hub SDK and a TypeScript state machine for a 576 × 288 display. The official simulator exercises the SDK interface, native rendering, gestures, and simulated G2 microphone path.

Deepgram Nova-3 handles streaming speech recognition from 16 kHz PCM microphone audio. Provider keys stay server-side, and the client receives a short-lived transcription token. Domain-specific keyterms help recognition without exposing speech-engine settings in the glasses interface.

A local resolver provides the immediate path for known people, organizations, and terminology. It does not wait for an LLM. Server-side StepFun step-5-preview adds structured understanding over finalized transcript windows. The extraction contract covers entities, commitments, questions, and pending profile updates, with strict checks for JSON structure, confidence, duplicate segments, and exact evidence sentences. Suggestions never silently rewrite the transcript or become confirmed facts.

StepFun also powers the website’s “Ask your meeting memory” chat. You can ask about open follow-ups, your latest meeting, or what you have discussed about a topic such as MCP. It uses the mock meeting and relationship records, and successful live StepFun inference has been verified. The model key stays on the server.

The technical console and relationship archive use React, Vite, and IndexedDB. They separate transcript evidence, suggestions, and confirmed records. G2 Meets does not retain raw audio. All checked-in employees, companies, documents, and conversations are fictional.

How I used Devin

Devin Desktop owned one bounded subsystem: the server-side StepFun extraction endpoint, its strict shared contract, validation and failure handling, and 22 mocked-fetch tests. Its work covered duplicate inputs, invalid JSON, timeouts, evidence mismatches, and unavailable credentials. I integrated the caller and human-review interface separately. The repository includes a handoff record of what Devin planned, implemented, debugged, and verified.

Challenges I ran into

Every pixel matters on glasses. I repeatedly simplified the interface: one lockable tooltip instead of a collection of pinned cards, continuous scrolling instead of unnecessary pages, and gestures that work without a keyboard or phone.

The official simulator also behaves differently from a normal browser. Correct bridge initialization and native gesture handling mattered more than polishing a duplicate browser simulation.

Another challenge was keeping live transcription, seeded context, and optional model enrichment honest. The product needs to remain useful when a provider is unavailable, without presenting a replay or fallback as a live model result.

Accomplishments I am proud of

The official simulator has sent real microphone audio to Deepgram, received finalized sentences, resolved a person and a company-specific acronym, saved a recap, and supported a user-confirmed People update.

The experience connects a small, focused glasses interface to a fuller desktop view of conversation and relationship records. It makes the boundary between what someone said and what AI suggests visible and testable.

What I learned

Useful AI on glasses is not about showing more information. It is about choosing a small amount of relevant context, making it easy to browse, and keeping the wearer in control.

Separating fast deterministic retrieval from slower model enrichment also makes failures easier to explain and keeps the core experience responsive.

What's next for G2 Meets

Complete physical G2 testing and refine the layout based on actual wearability. Evaluate extraction quality across more conversation domains, finish validating cross-device console mirroring on physical G2, and add opt-in knowledge connectors while keeping sources, private notes, and confirmed facts separate.

Current prototype status

Live Deepgram transcription is verified in the official simulator; a successful physical G2 run is not yet confirmed. Successful live StepFun inference is verified, and the website chat uses StepFun to answer questions over mock meeting and relationship records. Deployed cross-device mirroring and cloud backup are not claimed as operational. The local resolver remains available when enrichment is unavailable.

Workday: Future of Work

G2 Meets helps a new hire understand the company while taking part in real conversations. It retrieves trusted context about colleagues, terminology, and earlier meetings, then proposes evidence-backed follow-ups for review. The demo uses fictional employees and internal documents, with sources and uncertainty visible in the console. No Workday API integration is claimed. The prototype does not evaluate employees or make employment decisions.

Built With

• deepgram
• devin
• even-hub-sdk
• indexeddb
• node.js
• react
• stepfun
• typescript
• vercel
• vite
• vitest