IkoAgent

IkoAgent is how we at Ikosagon showcase call-flow mastery as a product — not a free-chat bot.

Years of contact-center instinct live in the graph: stages, must-say lines, escalation gates, payment-plan rules, quiet-hour awareness, and clean exits. When a model key is present, Groq runs real NLP turns inside each node (intent + slots + natural reply). The graph still owns transitions and must-say compliance lines. Keyword matching is the offline fallback.

For employers

• Authored flows as the artifact: versioned JSON skins for customer service, collections, and sales on one shared runtime.
• NLP-in-node: one Groq request per turn classifies the intent, fills slots, and drafts a natural reply for the node the graph predicts. The graph then decides for real; if the call lands somewhere else, the scripted line plays instead of a second request. Compliance nodes (allowParaphrase: false) lead with the must-say line verbatim. Keyword fallback when offline.
• RAG over a substantial policy/script corpus per vertical (returns, shipping, billing, hardship, quiet hours, ethical sales, and more); citations surface when the node asks for them.
• Mock CRM tools: lookup account, balance, payment plan (policy-bounded), promise-to-pay, case, callback, disposition — deterministic demo accounts (1001, 2044, 3300).
• Collections depth: plan negotiation inside policy limits, promise-to-pay with date capture and partial payments, settlement boundary, hardship escalation, dispute / already-paid reviews, cease and attorney requests, and third-party calls with no disclosure.
• Sales depth: needs discovery, corpus-backed plan fit, price / competitor / timing / approval objections, demo booking, and honest no-pressure upsell. Customer service adds cancellation with one save attempt, address changes, and repeat-caller de-escalation.
• Systematic evals: an offline regression suite (mocked model outputs, deterministic graph) covers routing, slots, identity, safety guards, verbatim compliance, and close routing, and runs before every merge. Results are on the eval report.

For clients

A controllable agent that can run CS / collections / sales workflows you approve — disclosures that cannot drift, escalation when the graph says so, and answers grounded in your procedures. We shape the experience with you; the model stays on a leash.

Honesty

• Interactive demo on this page: talk or type as the caller. No live inbound phone number (PSTN costs money; Hobby stays free).
• Browser TTS (Web Speech API speechSynthesis) optionally reads agent replies aloud — free system voices, mute toggle in the demo chrome. This is not production contact-center voice (no Twilio/ElevenLabs).
• Browser speech input (Web Speech API SpeechRecognition): hold or tap the mic to talk as the caller. Only the recognized text reaches our server — no audio. Chrome and Edge run recognition in the browser vendor's cloud; Safari support is partial; Firefox has no speech input, so typing stays the fallback.
• Conversational turns use Groq (CHAT_MODEL / openai/gpt-oss-20b) when GROQ_API_KEY is set; without a key the demo still advances on scripted lines + keyword intents. Rate-limited to protect Hobby quotas.
• Status is in progress: shared skeleton + rich corpus + live LLM turns in-node; telephony certification is not claimed.

How to try it

1. Pick a vertical skin, then a scenario card (or start a blank call).
2. Talk (hold the mic) or type as the caller. The agent desk shows the current step, collected details (accounts masked), the policy line cited, tool events and the outcome; the flow map lights the current step and traces the path. Leave Agent voice on to hear browser TTS, or mute it anytime.
3. Ask policy questions (returns, hardship, plan fit) to see RAG citations.
4. Tools fire as the graph allows. Restart anytime — turn cap is 20 per session.

Talk to us about a call-flow desk →