Technical Architecture

RAG Pipeline

sequenceDiagram
    participant V as Visitor
    participant M as Microphone
    participant W as Whisper ASR
    participant Q as Qdrant Vector DB
    participant G as GPT-4o (Groq)
    participant E as ElevenLabs TTS
    participant A as Avatar Display

    V->>M: Speaks question
    M->>W: Audio stream
    W->>Q: Query embedding
    Q->>G: Retrieved context + query
    G->>E: Response text
    E->>A: Synthesized speech + morph
    A->>V: Persona response

Performance Metrics

LayerComponentSpecLatency (ms)
CaptureClearOne BMA-CT12° beam, −48 dB noise floor<10
VisionIntel RealSense D43590 fps depth<10
ASROpenAI Whisper-tinyLocal processing120
InferenceGPT-4o 32k on Groq LPURAG prompt250–400
TTSElevenLabs22050 Hz synthesis350
Avatar mouthHeyGen live-talkReal-time sync60
Totaltarget < 1.5s

System Workflow

1. Input Selection

Visitors select a persona (Hoosier Oracle, Vonnegut, etc.) which triggers the idle state → transition animation → persona state sequence.

2. Voice Processing Pipeline

The system transcribes visitor speech, processes the query through GPT-4o (or other LLM), then converts the response to text-to-speech.

3. Real-Time Animation

Audio streams feed into facial animation systems (Audio2Face or Speech Graphics SGX) to generate lip-synced movement in real time.

4. Visual Stylization

Raw facial animation is transformed through a stylization layer:

5. Live Display Output

The stylized video with alpha channel is:

Voice Synthesis & Persona Morphing

Our AI system uses ElevenLabs’ voice cloning technology to recreate the authentic speech patterns of each Indiana persona. Historical recordings, when available, are used to train persona-specific voice models. For figures without recorded speech, we employ linguistic analysis of their written works combined with regional accent modeling to create plausible voice representations.

Fallback Strategy

Latency Management: Target response under 1.5 seconds during typical load. When processing exceeds this threshold, particle animation loops maintain engagement while displaying “The Oracle is considering…” status. Local fallback models activate during cloud service interruptions.

Oracle State Machine

The Oracle operates through a carefully orchestrated state system designed for smooth visitor interaction:

stateDiagram-v2
    [*] --> IDLE
    IDLE --> SUMMON : Motion Detected
    SUMMON --> LIVE_TALK : Voice Input
    LIVE_TALK --> LIVE_TALK : Conversation
    LIVE_TALK --> OUTRO : Silence (10s)
    LIVE_TALK --> GLITCH : Latency > 700ms
    GLITCH --> LIVE_TALK : Recovery
    OUTRO --> IDLE : Complete
    SUMMON --> IDLE : Timeout (30s)

State Descriptions:

Content Pipeline Architecture

Real-Time Rendering Stack:

  1. Audio Input → Whisper ASR → conversation state
  2. LLM Processing → character-specific response generation
  3. TTS Synthesis → ElevenLabs voice cloning with persona parameters
  4. Visual Composition → TouchDesigner/Unreal Engine real-time layering
  5. Display Output → Calibrated projection or holographic display

Visual Layer Management:

Failsafe Systems: When cloud services lag, the system transitions to cached responses with “thinking” animations until connectivity restores. Local TTS provides basic functionality during extended outages.

Holographic Display Technology

Pepper’s Ghost Implementation

Our primary display approach uses Pepper’s Ghost illusion technology, creating convincing 3D presence without requiring special glasses:

Alternative Display Options

Light-Field Technology: Looking Glass displays provide true parallax viewing for multiple simultaneous viewers (32”-65” units, $15K-$45K hardware cost).

Transparent OLED: Layered glass configuration allows see-through effects with particle depth layers (55” FHD panels, ~$16K per unit).

Volumetric Projection: Proto hologram units for reliable 3D presence in full installations.

Hardware Specifications

Core Computing

Display Infrastructure

Technical Documentation

Download the complete Bill of Materials and technical specifications for implementation.

📄 Download BOM.xlsx

Visual Design Language

Promotional Visual Language — Art Deco meets digital futurism

While the Oracle entities will appear as monochromatic, spectral forms — smoky particles and digital glitches bridging past and future — the promotional materials work in a related but distinct style. They lean on Art Deco geometry, vintage electrical motifs, and kinetic motion typography to shape an “old-meets-new” aesthetic for campaign pieces and adjacent projects.

Letterforms assemble and dismantle with machine-like precision, a gesture toward temporal echoes and a nod to Indiana’s manufacturing and craft traditions — from RCA’s years in Bloomington to the makers and tradespeople of the Showers district. The imagery borrows from moments when technology felt theatrical and full of promise — the polish of Deco, the spark of early electronics, and the speculative tone of classic sci-fi — and ties that sensibility to the state’s universities, notably IU in Bloomington, as well as Notre Dame and other research centers. Together, these elements create a visual language that connects industrial history, design heritage, and a sense of forward momentum.

Risk Assessment & Mitigation

Technical Risks

Operational Risks

Quality Assurance