# We Built an AI Employee Platform With Real Security - And Our AI Receptionist Just Answered Her First Phone Call

**TL;DR:** Atlas UX is a platform where 20+ AI agents work as employees - sending emails, managing CRM, publishing content, running intel briefs. We built enterprise-grade security first: tamper-evident audit chains, approval workflows, daily action caps, and a governance language that constrains what AI can do. Today, our AI receptionist Lucy answered her first real phone call. Here's how it works.

---

## The Security Stack

**Audit Chain:** Every mutation gets written to an append-only audit log with cryptographic hash chaining. Each entry's hash includes the previous entry's hash - tamper with one record and the chain breaks. Actor type, action, entity references, timestamps, IP addresses, full metadata. Nothing disappears.

**Decision Memos:** Agents can't approve their own risky actions. Spending above limits, recurring charges, or risk tier 2+ actions require a decision memo - what, why, cost, risk assessment, alternatives - that sits in a queue for human approval.

**SGL (System Governance Language):** A custom DSL that defines rules every agent follows. Action caps per day, spend limits, content policies, escalation rules. The engine evaluates SGL constraints at runtime before any action executes. Violations are blocked and logged.

**Engine Loop:** Orchestration ticks every 5 seconds, dispatching jobs with confidence thresholds. High confidence + low risk = autonomous. Low confidence or high risk = human in the loop.

---

## Lucy: AI Receptionist

Real inbound calls on a real number, routed through Twilio, processed in real-time.

**Call Flow:** Phone rings -> Twilio webhook -> bidirectional WebSocket -> decode mu-law to LINEAR16 PCM, upsample 8kHz to 16kHz -> Google STT (streaming, speaker diarization) -> Lucy's reasoning engine -> Google TTS (Neural2-F voice) -> downsample, encode mu-law -> caller hears Lucy speak. Round trip: 2-3 seconds.

**Caller Classification:** Every few exchanges, Lucy evaluates caller type (warm lead, VC, frustrated customer), sentiment (-1.0 to +1.0), energy level, and conversation mode. This adapts her behavior in real-time. VCs get composure and data. Frustrated callers get acknowledgment first.

**ContextRing:** Lucy can be on a Zoom meeting AND answering a phone call simultaneously. Both instances share memory through an in-memory state holding transcripts, speaker maps, and conversation context. Phone Lucy and Zoom Lucy are the same brain.

**Post-Call:** Generates summary with action items, saves MeetingNote, creates CRM contact activity, writes audit log, captures new leads, posts to Slack. All audited.

---

## Try to Stump Her

Call Lucy: **573.742.2028**

- Grill her on unit economics as a VC
- Be frustrated and see if she de-escalates
- Ask something deeply technical
- Switch topics mid-sentence
- Be rude

Real Twilio, real Google STT/TTS, real LLM reasoning, real CRM lookup, real Slack alerts. Every call classified, summarized, and logged.

**Come back and tell us what happened.** We review every transcript. Your call makes her better.

---

**Stack:** Fastify 5 + TypeScript, PostgreSQL/Prisma, Google Cloud STT/TTS, Twilio Media Streams, multi-provider LLM routing (OpenAI, DeepSeek, Cerebras), React 18 + Vite, Electron desktop app.

*Built by one person and 20+ AI agents. Not raising - building.* **[atlasux.cloud](https://atlasux.cloud)**

510 E Washington Street, Vandalia, Mo 63382

Atlas UX, a standalone multi-platform ai employee who works where you work

This release was published on openPR.