🌱 English documentation
日本語 →

🌱Getting Started — The Big Picture

For first-time readers. How calls are protected, where the voice comes from, and where money is spent — in four diagrams.

Aegis is a gatekeeper for your company’s phone line. Nuisance calls are stopped at the door. Important calls ring through to a person as usual. Only the calls in between are answered by the AI receptionist (the AI layer — an LLM), which asks what the caller wants.

This page is a guide for first-time readers. Technical terms appear as “friendly name (formal name)” pairs. Four diagrams explain the overall flow, where the voice comes from, where money is spent, and why it gets cheaper over time.

Diagram 1: The overall flow

📞 An incoming call 🚪 Gatekeeper (number check) just matches lists — ¥0 On the block list a known nuisance number 🔊 Recorded announcement pre-made audio — ¥0 Hang up the AI never wakes up — ¥0 Someone we know clients, family (whitelist) 👤 Straight to a person rings as usual — ¥0 An unknown number only here does the AI wake up 🤖 The AI answers asks the business — only paid part Decision should this call go through? To a person Hang up
Diagram: every call first meets the gatekeeper, which only matches the number against lists (free). Numbers on the block list get a pre-recorded announcement and are hung up; known callers ring straight through to a person. Only unknown numbers wake the AI receptionist — the one and only paid part — which talks, decides, and either passes the call to a person or hangs up.

There are only three things to remember.

  • Every call first meets the gatekeeper (the rules layer), which just matches the number against the lists (block list / whitelist) — so it is free.
  • Numbers on the reject list (block list) get a recorded announcement and are hung up. Callers on the trusted list (whitelist) ring straight through. Both are free.
  • Only for unknown numbers does the AI receptionist wake up and ask the caller’s business. That is the only paid part.

Diagram 2: Where the voice comes from

When blocking 🔊 Play a recorded file made in advance with VOICEVOX cost per call: ¥0 When the AI answers OpenAI Realtime (one service) 👂 Ears (listening) 🗣️ Mouth (speaking) 📝 Transcription no separate STT/TTS is ever called 💡 Swapping the decision brain (rules / LLM) does not change where the voice comes from
Diagram: two voice sources. When blocking, Asterisk just plays an announcement file pre-generated with VOICEVOX, so each call costs nothing. When the AI answers, ears, mouth, and transcription are all inside the single OpenAI Realtime service — no separate speech-to-text or text-to-speech is called. The decision brain can be swapped without touching either voice source.

The “blocking voice” and the “AI voice” come from different places.

  • Block announcements are pre-made recordings (generated in advance with VOICEVOX). Playing a file costs nothing per call.
  • When the AI receptionist talks, one service — OpenAI Realtime — does everything: ears, mouth, and transcription. No separate speech services are bolted on. The voice travels between the phone and the AI through the bridge (bridge.py — the voice relay).
ℹ️The brain and the voice are separate

Even if the decision brain — the gatekeeper (rules layer) or the AI receptionist (AI layer) — is swapped out in the future, the voice sources stay the same. Because they are separate, each part can improve on its own.

Diagram 3: Where money is spent

Incoming Gate check ¥0 During the call (3 patterns) Call ends Record ¥0 block Recorded announcement → hang up — ¥0 human Straight to a person — ¥0 ai The AI talks — paid only while connected to Realtime ¥ ・The LLM decision (classify_call) is included in the Realtime session = no extra decision fee ・The rules-layer decision is always ¥0 ※ ¥0 / paid here means AI usage fees (OpenAI API) only. Phone-line call charges are not included (they depend on your line contract).
Diagram: the timeline of one call. Incoming call → gate check (¥0) → during the call, block (recorded announcement, ¥0) or human (straight through, ¥0) or ai (paid only while connected to OpenAI Realtime) → call ends → record (¥0). The LLM decision runs inside the Realtime session, so it adds no extra fee, and the rules-layer decision is always free. ¥0 / paid here refers to AI usage fees (OpenAI API) only — phone-line call charges depend on your line contract and are not included.

The answer is simple: you pay only while the AI receptionist is actually talking. Everything else — the gate check, playing recordings, ringing a person, saving the record (call log) — is free.

The in-call LLM decision (“is this call a nuisance?”) runs inside the same Realtime session as the conversation, so it adds no extra fee.

Note that “¥0 / paid” in the diagram refers to AI usage fees (OpenAI API). Phone-line call charges depend on your line contract and are not included here.

Diagram 4: It gets cheaper the more you use it

📒 Lists grow learning from every call ❓ Fewer unknowns more calls are pre-sorted 🤖 AI wakes up less the paid part shrinks 💰 Costs go down the more you use it the more it is used, the smarter and cheaper it gets 🔭 Future: distillation lets the brain graduate too (→ One Step Ahead)
Diagram: a positive loop. Every call teaches the lists, so fewer numbers are unknown; the gatekeeper pre-sorts more calls, the AI (the only paid part) wakes up less, and the bill goes down. In the future, distillation lets even the decision brain graduate from the cloud — see One Step Ahead.

Every call teaches the lists. As the lists grow, fewer numbers are “unknown,” and more calls are settled by the gatekeeper alone. That means the paid AI receptionist wakes up less and less, and the bill goes down.

One step further ahead, there is a plan to move the decision brain itself onto the on-site device (distillation). See One Step Ahead.

Where it connects (today and concepts)

  • Today: being developed for Rhodium (the Hikari Denwa home gateway).
  • By design: any environment that speaks Asterisk / SIP can connect. The bridge’s front door is the standard AudioSocket protocol, and the phone side needs a minimal five-line configuration to connect.
  • ⏳ Concept: expanding to business phone systems, cloud PBXs, and other phone services. ⏳ Concept stage Still at the concept stage — no implementation or testing yet.

For the sales-channel side of this (PBX OEM), see Strategy & Pricing.

Mini glossary

  • Rules layer — decides using fixed rules only (numbers, lists). The “gatekeeper” on this page.
  • AI layer (LLM) — an AI that decides while talking. The “AI receptionist” on this page.
  • Block list — the list of numbers to reject. Recorded announcement → hang up.
  • Whitelist — the list of trusted numbers. Straight through to a person.
  • Bridge — the relay that carries the voice between the phone and the AI (bridge.py).
  • Shadow mode — a practice mode that records what a new decision method would have done, without changing live behavior.

Full definitions are in the Glossary.

ℹ️Recommended next stops