Voice + Canvas Analytics Copilot
Talk to your marketing data. It renders the analytics on a canvas, live.
Fortune-500 semiconductor leader (synthetic-data demo)
The problem
Marketing analytics sit in static dashboards no one has time to read. Answering "where should the next dollar go?" means waiting on an analyst, and attribution is usually last-click and indefensible. Leaders need answers in the moment, in plain language, that hold up to a CFO.
What it's good for
- Marketing and revenue leaders who want answers by voice, instantly
- Teams that need defensible attribution (MMM + counterfactual), not last-click
- Exec readouts and board decks generated on the fly
- Anyone who wants an analyst that never sleeps and remembers context
Overview
A voice copilot for B2B marketing intelligence. You talk to it and it renders the right analytics on a canvas, builds dashboards on demand, maintains conversational context and user preferences, and orchestrates a team of analytic agents. Powered by Gemini 3.1 native-audio Live (voice in and out, with function-calling) and an iridescent WebGL orb that reacts to your voice. Behind it sit eight agents: Signal Scout (morning brief), Propensity Modeler, Attribution Analyst (Bayesian marketing-mix modeling with a counterfactual ledger), Competitor Watch (live web research), ROI Optimizer, Buying Group Analyzer, GEO Intelligence, and Measurement Forecaster. Preferences persist as memory and ripple into agent behavior. Actions (email, agency brief, deck export) are always staged for human approval, never autonomous, and the whole thing runs offline on a deterministic scripted rail when no API key is present.
Key features
Voice copilot
Gemini 3.1 native-audio Live: speak naturally, it speaks back, and it calls tools to render views, edit charts, and run research. A WebGL orb swells with your voice.
Live canvas
Charts and dashboards render on command, on the fly: funnels, attribution forest plots, response curves, saturation, buying-group heat grids.
Eight analytic agents
A configurable team that runs overnight and produces insights; changing an agent knob recomputes the analysis and ripples into the KPI cards.
Persistent memory
Voice-stated preferences ("we cut programmatic this quarter") are remembered and shape how the agents behave.
Counterfactual attribution
Bayesian MMM with a counterfactual ledger: "if we zeroed trade media, we would lose these design wins, worth this much."
Board-ready export
Generates a PPTX/PDF deck with the insights, an AI critic pass, and a CFO objection sheet.
How it works
- 1You speak; Gemini 3.1 Live transcribes and decides which tools to call.
- 2Tools (render_view, edit_canvas, show_insight, web_research, remember, configure_agent, act) execute against the analytics engine.
- 3The canvas updates live; insights are cached and re-ripple whenever an agent knob or a remembered preference changes.
- 4Competitor and GEO questions hit live Google Search grounding for real, current answers.
- 5When you ask to act, it stages an email, brief, or deck export for your approval, then a backend builds the artifact.
How to make it live
- Frontend: `npm run build` outputs a static dist/ (vite base is relative). Host on Vercel, Netlify, or Firebase Hosting.
- Backend: containerize with the included Dockerfile and deploy to GCP Cloud Run (handles deck/PDF export and email send).
- Set VITE_GEMINI_API_KEY for live voice; optionally GMAIL_USER + GMAIL_APP_PASSWORD for real email export.
- No key needed to demo: it runs fully offline on a deterministic scripted voice rail with cached research and a mock outbox.
cd demos/ti-voice-canvas npm install cp .env.example .env # add your keys (optional) npm run dev # http://localhost:5173
Tech stack
Why it matters
- Voice-driven canvas analytics with sub-second native audio
- Defensible attribution: Bayesian MMM + counterfactual ledger
- Board-ready decks generated and critiqued on the fly
Client anonymized for public display; all figures and accounts are synthetic demo data.
Want a walkthrough of this build?
Happy to demo it live and talk through the architecture.