OMA — Online Meeting Advisor

AI copilot that listens to your online meetings, live.

Inspiration

Every meeting tool today works on the transcript afterwards — summaries, action items, recaps. Useful, but the moment where it actually matters is already gone.

I wanted a coach that sits beside you during the conversation: quiet when nothing matters, sharp when something does, and configurable to what you're actually trying to achieve in this specific meeting.

What it does

OMA is a desktop app (Mac/Windows) that listens to your online meetings from both sides — your microphone is you, the call's system audio is them — transcribes speech on-device, and surfaces short insight cards in real time: each one a headline, a priority, and a line of context. Because the two channels are captured separately, OMA always knows who said what. No recording leaves the machine; only distilled text snippets are analyzed.

Built with cost and quality in mind:

• A cheap gate (GPT 5.4 Mini) watches the transcript and decides "is there anything worth saying right now?" — most of the time the answer is no, and that costs almost nothing.
• Only when the gate fires does the AI coach (GPT 5.6 Sol, medium) run, producing a typed, prioritized card.

Before the meeting you paste in context — agenda, goals, background — and the app distills it into a brief the engine reasons against.

Coach Profiles

Insight Radar

Surfaces the interesting stuff as it emerges: hidden assumptions, contradictions, decisions being made implicitly, risks nobody named, connections between things said twenty minutes apart.

Goal Keeper

You define what this meeting must accomplish. The coach tracks the conversation against those goals and speaks up when one is being reached, skipped, or drifting.

Message Delivery

You have key messages that need to land. Because OMA hears you and the other side on separate channels, it knows whether you actually made a point and how they responded.

Build your own

Build your own profile

Every profile has a simple frequency dial — Quiet / Balanced / Chatty — that trades card volume against focus and API cost.

How it's built

• Desktop app (Mac/Windows) with a dual on-device audio pipeline: microphone (you) + system-audio loopback (the call) → voice-activity segmentation → Whisper transcription, all local
• GPT two-tier engine — GPT 5.4 Mini as the salience gate, GPT 5.6 Sol for insights, structured output via tool schemas, prompt caching to keep the frequent gate calls cheap
• Live browser UI streamed over Server-Sent Events: insight cards, a real-time activity feed of what the coach is thinking, pin & export
• Insight cards with priorities (minor / major / critical), deduplication, cooldowns, and adaptive pacing so the coach never becomes noise

Built With

• desktop
• javascript
• openai
• openrouter
• python
• whisper