Safebots listens to your conversation and puts real-time fact checks, charts, profile cards, and archive clips on screen — without interrupting you. You approve. It shows. No one knows how you did it.
You didn't touch anything. You didn't stop talking. Your guest didn't notice. But the audience — on Zoom, on your livestream, on their phones — just saw a sourced, animated stat card appear on screen with the verified number, the year-over-year delta, and the Bloomberg citation.
That's Safebots. A live AI intelligence layer that watches your conversation, proposes visualizations, and only puts them on screen after you approve — silently, without anyone knowing you're involved.
Open the URL on your phone. The shared screen follows every scroll, every tap, every voice command. No app. No Bluetooth. No cables. Just a URL and a WebSocket.
Scroll a PDF — the projected screen scrolls identically. Advance a slide — 500 participant phones update in the same frame. Pause a video — it pauses everywhere. The audience follows on their phones too.
AI detects it, verifies it against live web data, animates a stat card with source and delta. In the time it takes to finish the sentence.
Profile card: bio, handle, topic tags. Pulled from your graph if they're in it. Searched from public data if they're not.
"LIDAR." "ZK-proof." "Federated learning." Plain-English definition appears, calibrated to the conversation's evident expertise level.
The AI identifies it, surfaces headline, key claim, direct link. Participants on their phones can tap through immediately.
Comparison card. Both sides. Sourced from public data. You don't need to have memorized it.
If you have access to Bob's location in the graph, a map appears. If the other person tries the same query, they see a private error. Only you see the map.
Before anything appears on the shared screen — the Zoom, the livestream, the projector — it appears on your private dashboard with a countdown ring. Say "no" or "skip." Say nothing and it commits. The audience sees only what you approve.
"The host appears effortlessly informed. The audience sees nothing unusual. The AI did the research; the host did the judgment."
You also get private coaching the audience never sees. "This contradicts what the guest said in March." "This figure is lower than what they published last quarter." Sourced from your archive. Delivered only to your screen. The audience has no idea why you look so prepared.
Each person in the conversation streams their own microphone over their own authenticated connection. The server knows whose voice it is from the connection — not from trying to figure it out acoustically. So permissions are cryptographic, not probabilistic.
The AI pipeline never sees data the requester doesn't have permission to access. The access control isn't a prompt instruction the model could decide to ignore — it's enforced at the data layer before anything reaches the model. Bob never gave Robert access to his home address. The system reflects that fact structurally.
The same applies to navigation commands. Voice commands from a guest don't advance your slide. "Next slide" spoken by you advances the slide. Spoken by your guest, it's ignored — silently, with no visible awkwardness.
Every layer is independently useful. The AI enhances what's already there. None of it requires the layer above it.
| Capability | No AI | With AI (public data) | With AI + internal data |
|---|---|---|---|
| Phone as clicker — scroll PDFs, advance slides | ✓ Works now | ✓ | ✓ |
| Webpage sync — scroll a website on your phone, shared screen follows | ✓ Works now | ✓ | ✓ |
| Live polling with QR code — audience answers on phones | ✓ Works now | ✓ | ✓ |
| Audience follows along on their phones in real time | ✓ Works now | ✓ | ✓ |
| B-roll gallery — images matching conversation keywords | ✓ Browser speech, no API key | ✓ More precise queries | ✓ |
| Voice navigation — "next slide", "zoom in", "highlight the top bar" | ✓ Pattern matching, <10ms | ✓ | ✓ |
| Real-time CSS theming — change colors mid-show | ✓ Works now | ✓ AI detects register shifts | ✓ |
| Fact verification — numbers checked and sourced | — | ✓ Web search | ✓ |
| Profile cards — bio, handle, tags for named people | — | ✓ Public data | ✓ Internal graph |
| Private coaching — contradictions, corrections, context | — | ~ Limited | ✓ Full archive |
| Access-controlled data — different content per viewer | — | — | ✓ Substrate enforced |
| Directions to "Bob's house" — resolves who Bob is and if you have access | — | — | ✓ Per speaker identity |
| Audit trail — what appeared, when, sourced how, approved by whom | — | — | ✓ Cryptographic |
A Safebox workflow has already ingested your channel — every episode, every video, every audio file. It transcribed the audio, indexed it by topic and keyword, and cut it into clips. Now the AI knows what you've covered before, and can surface exactly the right moment from your own archive while you're live.
A Safebox workflow ingests your channel in the background — transcribing audio, extracting topics, cutting clips at natural break points. Done once. Available forever.
"Attention heads" appears in the live transcript. The archive returns matching clips instantly — pure keyword match, no model call, no latency. Works offline.
With the LLM active, "scaling laws for transformers" finds the clip where you discussed Chinchilla even if neither phrase appeared verbatim. The keyword index is the fast path; the LLM is the smart path.
The AI detects when the live conversation touches something you've covered before. The clip appears in your private veto queue with the relevant timestamp and transcript excerpt. You decide whether to play it.
Tap "Show on screen." The clip starts playing on the shared display — with the relevant section already cued. Tap "Private" to send it to a specific participant. Tap "Skip" and it's gone.
You say it. The AI finds it. The clip appears. You play it. The audience sees continuity across your entire back catalog. No producer needed. No prep required.
"The guest says something you covered six months ago. The clip appears in your queue at the exact timestamp. You tap play. The audience thinks you have a perfect memory."
| Archive capability | How it works | Requires LLM? |
|---|---|---|
| Find clips by keyword from live transcript | Inverted index, sub-100ms lookup | No |
| Find clips by semantic meaning | Embedding similarity search | Yes |
| Detect when current conversation matches prior episode | Keyword overlap + topic graph | Optional |
| Surface clip with correct timestamp cued | Clip boundary metadata from ingest | No |
| Transcribe new episodes automatically | Safebox ingest workflow (Groq Whisper) | No |
| Surface contradictions between guest's claim now vs 6 months ago | LLM cross-references session entity graph with archive | Yes |
| Access control — some clips private to host only | Stream-level permissions, substrate enforced | No |
When the conversation touches a trend, a comparison, or a claim — Safebots can pull the data and render a chart on the shared screen in seconds. From public sources, from an ingested knowledge base, or from private data the requester has access to. The chart appears. The host decides whether the audience sees it.
Guest mentions a funding figure. Safebots queries public data, renders this.
From your ingested archive of tracked sources — not on the public internet.
From a tracked list of researchers ingested into your knowledge base — not a public API call.
Blends public search data with your privately ingested source archive.
These charts are generated live. The host says "show me AI investment by year" — or the AI detects the topic from the conversation and proposes the chart without being asked. The host approves. It appears on screen.
Web search grabs the current figure, structures it into chart data, renders it in under three seconds. Bloomberg, Statista, research papers — wherever the data lives publicly.
A Safebots workflow has already indexed the outlets you track — Aligned News, specific newsletters, RSS feeds. Trend data from your archive that isn't anywhere on the public internet.
Twitter/X lists, LinkedIn signals, community channels — ingested on a schedule into your private knowledge base. Sentiment trends across a tracked researcher list. Topics your community is discussing this week versus last month.
Internal CRM, subscriber data, proprietary research — anything your organization has ingested and access-controlled. The chart appears only if the requester's identity has read access to the underlying streams.
One chart, multiple sources. Public web data for the industry baseline, your private archive for your own coverage, ingested social for community signal — all normalized and combined.
"Show me AI funding trends" — spoken or typed quietly while your guest is talking. The chart proposal arrives in your veto queue. You approve with one tap. The audience sees a sourced chart that appeared from nowhere.
"The guest mentions a figure. You say nothing. A sourced bar chart appears on screen behind you. The audience assumes you prepared it. You didn't — Safebots did, ten seconds ago."
The control panel is the same chat the AI uses to show proposals. You type a query mid-conversation — quietly, while your guest is talking — and a card proposal comes back. You approve it. It appears. No one knows you asked.
Type "what's Anthropic's current valuation" while your guest is mid-sentence. A stat card appears in your veto queue 2 seconds later. Commit it or don't.
"Next slide." "Zoom in." "Highlight the top bar." These resolve in under 10ms from a vocabulary file — no model call, no latency.
When the AI finds three angles on a comparison, it posts them as tappable buttons. You pick one. That card gets built and committed.
Show the presentation alongside the intelligence feed on your Zoom or livestream. Viewers see the AI proposals, the host approvals, the access errors. Governance as theatre.
"Robert is the host. Greg is the guest. You say something. A sourced card appears. You try to look up something private. It says no. The audience sees all of it. No one planned any of it."
— The Safebots demo, May 25 2026Two screens side by side. Left: the presentation — b-roll running, cards floating over it. Right: the intelligence feed — proposals arriving, Robert approving some and discarding others, private coaching appearing only for Robert, access errors appearing only for Greg.
The moment that lands: Greg says "give me directions to Bob's house." An error appears in his chat — Bob never shared his address with Greg. Robert says the same words. A map appears on the shared screen. Same words. Different identity. Different outcome. No one wrote a rule about it. The data structure decided.