/LAB/EVEN-G2PORTFOLIO ↗

SIDE PROJECT · 05

EvenG2Glasses

Status · Shipped to a physical device

Year

2026

Role

Solo build

Focus

  • Smart glasses
  • Voice AI
  • Full-stack

(01) Context

The Even Realities G2 is a beautifully constrained device — a 1-bit, six-line monochrome display, a microphone, and a ring gesture, and that's it. I wanted to find out what genuinely useful AI looks like on a surface that small, so I built a whole suite of apps for it: speak a reminder, ask a question grounded in your own documents, get a live reply suggestion mid-conversation — all as glanceable text, in under a second.

The problem

Every assumption a normal app makes is gone: no keyboard, no scrolling, no color, almost no screen. The hard part wasn't the AI — it was making reasoning models feel instant and legible inside 240 characters of monochrome text, on real hardware, without dropping a tap.

(02) Approach

  • 01

    A suite, not a demo

    Five app modules over one platform: Loop (hands-free voice reminders), Recall (retrieval-augmented Q&A over your own PDFs and docs), Cyrano (real-time conversation suggestions), Capture (voice notes auto-structured by an LLM), and a glanceable trading HUD over a paper-trading system.

  • 02

    Provider-abstracted AI

    A single Claude interface with natural-language-to-structured-JSON parsing, RAG (chunking, embeddings, a SQLite vector store, persona system), SSE streaming, prompt caching, and a fast-model tier — plus real-time speech-to-text behind an intent router that decides which app you're talking to.

  • 03

    Full-stack, both ends typed

    A FastAPI backend (~8K LOC) with HMAC auth, upload caps, and decompression-bomb guards, and a TypeScript on-device runtime (~12K LOC) driving the 240-character, six-line, 1-bit display. OpenAPI-generated shared types keep the client and server from drifting.

  • 04

    Constrained-device UX

    Diagnosed and fixed real input-latency and dropped-tap bugs on the hardware, then rebuilt the flagship around a single-button, optimistically-rendered interaction so it feels instant even while the model is still thinking.

  • 05

    Infrastructure that heals itself

    Five self-healing launchd services, an HTTPS Cloudflare named tunnel stood up without disrupting a separate live tunnel on the same account, and a security-locked Telegram control bot — shipped to a physical device through the Even Hub developer portal.

(03) Outcome

A working, on-device AI platform — five app modules, five production services, ~25K lines — installed on real glasses, with 332 tests across 23 suites and an adversarial review that caught a timestamp-comparison bug before it shipped.

  • ~25K

    lines, full-stack

  • 332

    tests / 23 suites

  • On-device

    shipped via Even Hub