[004]·apr_2026v0.9

Flux

Voice AI
Gemini
Workflow JSON
12+ integrations

Voice-first automation platform on Meta Ray-Ban glasses — one spoken command triggers multi-step workflows across 12+ services, including Gmail, Slack, Notion, and Fetch.ai Agentverse. Built as BarelyAtWork at LA Hacks 2026, in a team of four over 36 hours.

Code ↗

flux — voice_to_workflow●

voice_transcript · ray-ban

"send yesterday's demo clips to the team and block 30 minutes tomorrow morning"

● captured

pipeline · gemini

✓transcribe42 ms
✓parse intent
✓generate workflow
✓validate paramscore loop
↻repair missing fields2 fixed
●readyexec

workflow.json · output

{
  "workflow": "share_demo",
  "trigger": "voice",
  "steps": [
    { "app": "gmail",
      "action": "send",
      "params_valid": true },
    { "app": "slack",
      "action": "post",
      "params_valid": true },
    { "app": "calendar",
      "action": "block",
      "params_valid": true }
  ],
  "status": "ready"
}

✓ validate→✓ repair→● ready3 steps · 12 integrations

InputRay-Ban voice command

ModelGemini constrained parser

Loopvalidate → repair

OutputExecutable workflow JSON

IntegrationsGmail / Slack / Notion / Calendar / Fetch.ai

How It Works

4 stages

S.01capture

Voice Capture

User speaks a messy natural-language command through wearable glasses.

S.02parse

Intent Parsing

Gemini extracts the workflow goal, target apps, actions, and required parameters.

S.03loop

Validation + Repair

A constrained schema checks missing or invalid fields, then repairs the workflow before execution.

S.04execute

Multi-App Execution

The final JSON triggers coordinated actions across email, messaging, calendar, and productivity tools.

Technical Decisions

2 tradeoffs

D.01

Structured JSON over freeform agents

Reliable workflows require predictable schemas, not open-ended model output.

D.02

Repair loop for messy speech

Voice commands are incomplete by nature, so the system fills and corrects parameters before running.