ANSIO · Conversational Alpha Console

A voice-native growth analyst for the creator economy. Ask out loud — ANSIO retrieves a live creator library through Moss in milliseconds and surfaces the underpriced creators, with every retrieval shown as it happens.

_{ANSIO — Artificial Narrative & Signal Intelligence Operating system}

---

🎬 Demo

Left: sessions · Centre: voice + text conversation · Right: the live Moss evidence stream with millisecond HUD. Talk → ANSIO scores the creator library and the recall chain lights up in real time.

▶️ To capture an animated walkthrough, see assets/RECORD_DEMO.md — record one voice turn on the running app and drop it in as assets/demo.gif.

---

🧩 The problem

Influencer pricing is a black box, and the market prices creators by follower count instead of real value:

• The same creator quotes wildly different rates to different brands; public rates are nearly impossible to scrape.
• Within one niche, cost-per-thousand-followers varies by orders of magnitude — the market is mispricing talent everywhere.
• Finding the genuinely undervalued creator means a human scrolling for a week, eyeballing follower counts that don't track real influence.

💡 What ANSIO does

ANSIO is a Bloomberg terminal for creators. A founder describes a growth goal in plain speech; ANSIO runs a real, multi-hop retrieval over a creator library and ranks candidates by alpha = real value ÷ asking price — the gap between what a creator is worth and what the market charges.

Underpriced ≠ cheap. Alpha rewards high value at a low price, not low price alone.

✨ Key features

🎙️ Voice-native	Talk or type — both reach the same agent. Real LiveKit WebRTC session with streaming captions that update in place, not a transcript box.
⚡ Moss real-time recall	A 5-hop retrieval chain measures ~9 ms end to end (P50 1.8 ms/hop, warm). A slow vector DB (~432 ms/hop) would stall the conversation for ~2 s.
🎬 Staged story reveal	Cards don't dump at once — the right rail advances beat by beat with the consultation (competitors → their playbook → similar-but-underpriced → alpha shortlist), enforced by an eval-tested tool-ordering discipline.
📊 Alpha ranking	Pre-computed 7-signal scoring; Moss only retrieves at runtime, so the answer is instant.
🧠 Consent-gated memory	Toggle Memory on and ANSIO distills each call into a user profile stored in Moss — next session it greets you knowing your product, platform and budget. Toggle off = zero reads, zero writes. One-click profile reset.
🗂️ Real session history	Sessions persist locally with their transcripts and their evidence streams — open any past consultation and the right rail replays its full card chain.
🌐 Language modes	EN / 中文 / Auto from settings — one language per reply (no mid-sentence mixing), wired through prompt, STT model and TTS voice.
🪟 Live evidence stream	Every retrieval pushes a designer-faithful card (sim-bars, donut, 4-column alpha leaderboard) with a real millisecond HUD plus a per-reply voice-latency badge — you watch Moss work.

🏗️ How it works

flowchart LR
    U([Founder voice / text]) -->|WebRTC| LK[LiveKit Agents 1.5]
    LK -->|STT deepgram/nova-3| AG[ANSIO agent]
    AG -->|LLM function calling| MM[LLM factory<br/>gpt-4.1-mini · MiniMax · Claude]
    AG -->|on_user_turn_completed<br/>+ recommend_kols meta-tool| MOSS[(Moss · local indexes)]
    MOSS -->|sub-10ms hits| SC[Alpha scoring<br/>pre-computed]
    SC -->|evidence cards| DC{{DataChannel}}
    DC --> UI[Three-panel console<br/>+ ms HUD]
    AG -->|TTS MiniMax| U

Loading

The staged reveal — the conversation is the demo. Each founder reaction advances exactly one step, and that step's retrieval paints its card on the right rail (an 8-beat evidence chain, mirroring a real consultation):

flowchart LR
    A["01 · competitor landscape"] --> B["02 · their creator playbook"]
    B --> C["03 · creator discovery"]
    C --> D["04 · audience intelligence"]
    D --> E["05 · creator alpha ranking"]
    E --> F["06 · bundle"]
    F --> G["07 · content"]
    G --> H["08 · ROI"]

Loading

Stage ordering is enforced two ways: a hard tool-ordering discipline in the system prompt (regression-tested by tests/staged_flow_eval.py against the live brain) and a per-turn card-type guard in the agent. Each hop is a Moss query at ~2 ms (P50 1.8 ms, measured warm); a full 5-hop recall chain is ~9 ms of retrieval. That speed is the product — at 432 ms/hop the agent would talk over a 2-second pause. This is why it has to be Moss.

🛠️ Built with

Tool	Role
Moss	Real-time semantic search — the multi-hop recall engine and the agentic memory store (user profiles). Cloud-first with a zero-quota on-device session-index fallback. Sub-10 ms, zero infra.
LiveKit Agents 1.5	Realtime voice pipeline (STT · turn detection · WebRTC transport + DataChannel).
MiniMax	Bilingual TTS voice; hot-swappable LLM option in the model factory.
LiveKit Inference gpt-4.1-mini	Default function-calling brain (env hot-swap: MiniMax-M2 / Claude).
Deepgram nova-3	Speech-to-text — multilingual model, with a dedicated Mandarin model when 中文 mode is selected.

🚀 Quickstart

git clone https://github.com/SkylarWJY/ANSIO-conversational.git
cd ANSIO-conversational

# 1. Backend deps (LiveKit worker + Moss tools + token server)
cd agent-py && uv sync && cd ..

# 2. Secrets — copy the template and fill in your own keys (never commit .env)
cp agent-py/.env.example agent-py/.env
#   LIVEKIT_URL · LIVEKIT_API_KEY · LIVEKIT_API_SECRET
#   MOSS_PROJECT_ID · MOSS_PROJECT_KEY · MINIMAX_API_KEY

# 3. Build the Moss indexes (generates synthetic + public-field data)
cd agent-py && uv run python src/build_indexes.py && cd ..

# 4. Run everything (single origin on :8788 — page + /token, plus the worker)
pm2 start ecosystem.config.cjs        # web + token + agent worker
# …or run the three processes manually (see ecosystem.config.cjs header)

open http://localhost:8788/            # grant the mic, then click "Talk to ANSIO"

One port (8788) serves the static console and the LiveKit token endpoint, so a single SSH tunnel (ssh -L 8788:localhost:8788 …) drives the whole demo — LiveKit media flows directly to the cloud, not through the tunnel.

📁 Project structure

ANSIO-conversational/
├── app/                 # three-panel voice console (vanilla JS, zero build)
│   ├── index.html       #   left rail · centre conversation · right evidence stream
│   ├── bridge.js        #   LiveKit web bridge (connect · mic · audio · DataChannel)
│   └── vendor/          #   livekit-client UMD
├── agent-py/            # LiveKit Agents worker
│   └── src/             #   agent.py · llm_factory · 5 tools + recommend_kols meta-tool
│                        #   memory (consented profiles) · lang (EN/中/Auto modes)
│                        #   moss_router (on-device fallback) · scoring · events
├── token_server.py      # FastAPI: signs LiveKit JWTs + serves the static site
├── ecosystem.config.cjs # pm2 process map (local dev)
└── demo/                # original landing demo

🔒 Data & confidentiality

ANSIO is built on public creator fields (name, handle, follower count, collaborated brands) plus a synthetic dataset. Any real deal pricing is collapsed into an aggregated benchmark only; individual quotes never enter the indexes, never appear on screen (the UI shows “Estimated Market Cost”), and raw data files are git-ignored. Credentials live only in a git-ignored .env. User memory is opt-in: profiles are written to Moss only while the Memory toggle is on; switched off, the agent performs zero profile reads or writes, and a one-click reset deletes the stored profile.

👥 Team

Built by a cross-border team for the YC × Moss Conversational AI Hackathon.

	Role
@baizhiyuan	Backend — LiveKit agent, Moss retrieval & recall chain, alpha scoring, token server
@SkylarWJY	Frontend — three-panel console design & demo
@clfhaha1234	Product — PRD & retrieval contract

_{Engineered with Claude Code.}

🙌 Acknowledgements

Built in 24 hours for the YC × Moss Conversational AI Hackathon. Thanks to Moss, LiveKit, MiniMax, Deepgram, and Y Combinator for the tools and the arena.

📄 License

MIT © 2026 ANSIO Team

🛰 More from Skylar

• skylarnyc.com — NYC date-spot guide: 100 review-backed picks, built with the open-source city-food-guide engine
• More projects → github.com/SkylarWJY