Kian Analytics
A bot that answers well is table stakes. A bot that predicts the next move, surfaces its own blind spots, and builds a custom seminar with you — that is what Kian does now.
This release is three shifts at once. First, analytics stopped being retrospective. A predictive intelligence layer scores every live session for lead probability and drop-off risk, surfaces next-best-action suggestions, and raises anomaly alerts when the booking rate deviates from its fourteen-day baseline. Second, conversation intelligence shows the shape of every session — intent trajectory from curiosity to decision, friction points where users drop off, knowledge gaps where Kian guessed instead of knowing, and a citation-accuracy meter for every RAG source. Third, and most important for WTM: the Custom Seminar Builder. Users walk through a four-step guided flow — team context, goals, format, brief — and Kian generates a structured PDF seminar outline that goes straight to the trainer team as a hot lead. No pricing anywhere in the flow, because WTM seminars are always custom-scoped.
Under the hood, the changes are systemic. A new Operator Studio consolidates prompt tuning, knowledge editing, A/B test control, and Connor pattern inspection into one workspace. Multi-modal input — image and PDF upload, push-to-talk voice, optional spoken replies — turns the widget into a working surface rather than a text box. Knowledge Intelligence runs a nightly gap-finder that nudges admins when a question has been asked repeatedly without a confident answer, and an auto-refresh job re-embeds seminar pages when their source content changes.
Predictive intelligence on every live session
Lead probability, drop-off risk, and next-best-action suggestions — computed after every bot turn, refreshed every five minutes.
The predict-session edge function runs after each assistant message. It combines Connor's learned patterns (intent, sentiment, engagement signals), chat analytics (messages, reactions, scroll velocity, idle time), and a fourteen-day baseline per company to score lead probability (0–100) and churn risk. Next-best-action is looked up from the top-performing pattern for the current intent cluster. A scheduled anomaly-detector compares today's booking and resolution rates to the baseline and raises alerts to email and the in-app bell when z-score exceeds two.
Conversation intelligence — every turn, visible
Intent trajectory, friction points, knowledge gaps, and a per-turn confidence heatmap for every session.
Each session now renders as a timeline: classification shifts from informational to comparative to decisional are drawn as a ribbon. Friction-point aggregation highlights which last-messages precede abandonment. Knowledge gaps — Kian answers with confidence below 0.6 — feed a dedicated admin queue with frequency counts. The confidence heatmap colours each turn by the classifier's certainty, making guesswork obvious at a glance.
Custom Seminar Builder — WTM flagship
A four-step guided flow that turns a conversation into a structured seminar brief. PDF out, trainer in.
Step one captures team context (industry, size, target audience). Step two captures goals via free text plus quick-chips (Kommunikation, Change, Führung im Homeoffice). Step three captures format preferences (Präsenz, Remote, Hybrid, Dauer). Step four generates an individual seminar outline — topic, modules, methodology, audience fit, next steps — rendered as a PDF via react-pdf and emailed to the WTM trainer team as a hot lead. No pricing at any point in the flow, because every WTM seminar is custom-scoped.
Multi-modal input: image, PDF, voice
Users upload screenshots or PDFs and Kian analyses them. Push-to-talk voice input with optional spoken replies.
Uploads hit /api/chat/upload with a 10MB limit, PII scan on ingest, and a thirty-day retention policy in a dedicated Supabase Storage bucket. Vision analysis uses Claude Sonnet 4.6 vision — the same model already in the stack. Voice input uses the Web Speech API with a Whisper fallback for browsers without it. Spoken replies are opt-in via a data-voice widget attribute and use OpenAI TTS. Upload and voice both count as strong lead signals and feed the lead-score model directly.
Operator Studio: prompts, knowledge, A/B in one place
Live prompt tuning with preview and rollback, a versioned knowledge editor, and visual A/B test control — no code, no redeploy.
The new /studio route replaces four separate admin flows. Prompt edits open a diff view against the current live version, a preview pane streams a test response, and rollback is one click. The knowledge editor uses markdown with metadata and keeps a version per edit. The A/B panel exposes traffic-split sliders, live KPI tiles per variant, and a winner-declare button that promotes the winning variant and retires the loser. Connor's learned patterns are inspectable — each one shows confidence, decay, and a toggle to disable without deleting.
Knowledge Intelligence: gaps, refresh, ROI
Nightly gap-finder, auto-refresh on seminar page changes, citation-accuracy per source, and knowledge-ROI attribution.
The gap-finder clusters low-confidence questions and surfaces clusters above frequency three to the admin queue with a one-click 'draft an entry' action. The auto-refresh cron polls the WTM sitemap, diffs each seminar page against its last-known hash, and re-embeds changed content. Citation-accuracy is measured per knowledge item — how often the source Kian cited actually contained the asserted fact. Knowledge-ROI ranks items by downstream conversion lift, so admins know which entries earn their keep.