How it works

From conversation
to intervention.

TiQ is a five-layer system. Each layer is designed for one purpose: to surface the right insight at the moment it can still change something.

The pipeline

Five stages. One purpose.

Every suggestion TiQ surfaces has traveled this entire pipeline — in under five seconds from when you last spoke.

// 01
Audio capture
Chrome extension captures both sides of the conversation — your microphone and the meeting tab audio — merged into a single stream.
Chrome Extension
→
// 02
Transcription
Deepgram Nova-3 converts audio to final transcripts in real time. Only confirmed utterances enter the reasoning engine — interim results are discarded.
Deepgram Nova-3
→
// 03
Conversation buffer
A 120-second sliding window accumulates final utterances. A hybrid trigger gate decides when there's enough new signal to call the reasoning engine.
In-memory only
→
// 04
LLM reasoning
The reasoning engine receives the transcript window plus session context, memory, and radar state. It returns structured suggestions and the current conversation state.
GPT-4o / Claude
→
// 05
Private delivery
Confidence-gated suggestions appear on your dashboard or Chrome side panel — only. Nothing is read aloud, nothing is visible to others. The side panel sits alongside your meeting tab so you never have to switch windows. A strategic whisper, not an announcement.
Your screen only
Before the meeting

The briefing that makes suggestions surgical.

TiQ asks seven questions before you start. The answers transform suggestions from interesting to precise — and set the boundaries that protect your credibility in the room.

1
Who are you in the room?
Your role shapes everything — which questions are relevant, which risks matter, what language to use. A software architect needs different prompts than a product leader.
→ Impact: Loads your persona configuration. Questions, risk types, and panel layout adapt to your role.
2
What kind of meeting is this?
An architecture review needs different intelligence than a stakeholder alignment or a client negotiation. Session type shapes the radar model and which suggestion types are prioritised.
→ Impact: Determines which conversation states TiQ watches for and which signal types are most relevant.
3
Who are you meeting with?
The relationship context is the single most important field. It tells TiQ what it is and isn't allowed to surface — protecting your credibility in the room. A new client meeting demands different intelligence than an internal review.
→ Options: Internal team · New client · Existing customer · Vendor / partner · Leadership · Interview
→ Impact: Sets a hard permission boundary on suggestion types. New client: questions and context only — no risk flags, no challenges to their organisation. Leadership: high-signal, crisp, decision-focused only.
4
Meeting brief — goal, decision, strategy, stakes
Four inputs that prime the reasoning engine. This is where TiQ stops being a transcript analyzer and starts being a strategic advisor. Takes ten seconds.
→ Goal: gives TiQ a compass. Suggestions evaluate whether the conversation is moving toward it.
→ Potential decision: TiQ watches for forming signals, missing checkpoints, and blockers.
→ Strategy: Persuade / Defend / Explore / Align / Inform / Negotiate — shifts which suggestion types are prioritized.
→ Stakes: Critical meetings get higher confidence thresholds. Fewer suggestions. Each one more precisely timed.
5
What should TiQ watch for?
Optional watch signals tell TiQ what topics matter most to you in this specific meeting. When those signals surface in the conversation, TiQ elevates them as strategic moments rather than passing remarks.
→ Options: Budget / pricing · Timeline pressure · Scope disagreement · Missing stakeholder · Security / compliance · Technical feasibility · Ownership gap · Political tension — and more.
→ Impact: Suggestions anchored to your watch signals surface with higher priority. The system detects when your flagged topic enters the conversation even when no one names it directly.
6
LLM provider and API key
Bring your own OpenAI or Anthropic key (BYOK), or use TiQ-managed access with no configuration. Your key is held in React component state only — never written to localStorage, sessionStorage, the server, or any log.
→ BYOK: Key lives in browser memory for session duration only. Cleared on session end. Unreachable by any external system.
7
Privacy acknowledgement and launch
One sentence confirming your conversation stays yours. Chrome is primed to ask tab audio permission. Everything is ready before you click Start.
→ On Start: Session token issued. WebSocket established. Audio capture begins. Dashboard goes live.
What TiQ surfaces

Four types of intelligence.
Each earned at a different threshold.

Questions to Ask
The highest-value panel. Surfaces questions the conversation hasn't answered — clarifying, probing, risk-oriented. Ranked by relevance. The question that could change the meeting outcome appears at the top.
Confidence ≥ 0.70
// Example
RANK_01 // SLA_BOUNDARY  ·  Ask
A deployment decision is forming but the SLA impact of the new service boundary hasn't been confirmed.
→ Has the inter-service latency been validated against the 99.9% SLA already committed to customers?
The insight above. The question you can say right now below.
Risk Flags
Unaddressed assumptions, gaps in coverage, decisions being made without essential information. Maximum three active flags at any time. Auto-dismisses when the topic is addressed in subsequent conversation.
Confidence ≥ 0.78
Decisions Log
Captures explicit decisions as they are made — with a session-relative timestamp. Tracks the forming and confirmation lifecycle. You walk out with a permanent record of everything decided, without taking a single note.
Confidence ≥ 0.80
Action Items
Commitment language detected in real time — who said they'd do what. An interactive checklist that you can check off during the session. Copy the full list with one click when the meeting ends.
Confidence ≥ 0.75
Conversation Radar

TiQ knows where the meeting
is heading, not just what was said.

On every reasoning call, TiQ classifies the conversation into one of eight states — and detects overlay conditions that cut across any state. This is what makes suggestions timely rather than merely accurate.

8 Base States
One unified model across all meeting types
Scanning Opening — participants orienting, no clear direction yet
Framing Problem space being defined — the agenda is being set
Exploring Options or approaches being considered — productive divergence
Decision forming Convergence happening — a choice is taking shape
Commitment Agreement reached — the decision is being locked in
Closing Next steps and owners being confirmed — meeting wrapping
Tension Interrupt — unresolved conflict or disagreement surfacing
Drifting Interrupt — conversation moving away from the meeting goal
3 Overlay Conditions
Cut across any base state — detected independently
Blocked Progress is stalled — a decision or action is stuck
Missing voice A key stakeholder has gone silent or hasn't been heard
Risk introduced A new risk or constraint has entered the conversation
Overlays appear as badges alongside the base state — e.g. "Decision forming · Risk introduced"
Quality gate

Silence is better
than a bad suggestion.

Every suggestion type has its own confidence threshold. Lower stakes allow more exploration. Higher stakes demand precision. Below threshold — nothing is surfaced.

Suggestion type Minimum confidence Rationale
Question
0.70
Questions are low-risk. Being slightly off is acceptable — the user can discard the question. Exploration has value.
Action Item
0.75
Commitments matter. A false action item creates confusion about who owns what. Higher bar than questions.
Risk Flag
0.78
False risk flags erode trust quickly. If TiQ cries wolf, users stop reading flags. High bar protects credibility.
Decision
0.80
Highest stakes. A wrong decision log entry causes confusion, disputes about what was agreed, trust breakdown.
🔒
Everything described above stays in memory.
The audio, the transcripts, the LLM prompts, the suggestions — none of it touches a database or log file. When your session ends, the pipeline clears. The full privacy architecture is documented separately.
Read the privacy architecture →
Ready to use it

See the pipeline
in your next meeting.

Start with a single session. No commitment. The whole pipeline, live, in your next high-stakes conversation.

Privacy architecture →