Architecture

1. Overview

RobotRoss combines customer-facing input channels, an orchestration layer, local and cloud model components, the Huenit robot arm, and studio recording/output systems. The ATF architecture page should describe that operational chain first, then show how the ATF wraps it into a live technical file.

2. System Flow

• Input channels: Shopify orders, Telegram or scripted commands, and the live voice showcase.
• Conversation and control: chat_ross.py handles interactive voice sessions, while bob_ross.py acts as the main orchestrator for write, draw, svg, sketch, check, and calibrate.
• Model path: Claude Haiku 4.5 is used for control/commerce-side reasoning, while Apertus 8B via Ollama is used for local narration and parsing tasks. Some voice/TTS flows also reference Kokoro, Voxtral, and system fallback engines depending on runtime mode.
• Execution layer: Huenit control scripts convert text, sketches, or SVGs into G-code streamed to the robot arm.
• Studio output: OBS records the drawing session, order data is written into the ledger, and customer-facing proof/output is produced from that recorded session.

3. RobotRoss Architecture Layers

Layer	Component	Description
Input	Shopify, Telegram, Voice Showcase	Human prompts, uploaded designs, or spoken requests enter through commerce, messaging, or live demo channels.
Control	chat_ross.py + bob_ross.py	Interactive session handling and the main job orchestrator decide what the robot should do next.
Models	Claude Haiku 4.5, Apertus 8B, Whisper, Kokoro/Voxtral	Planning, narration, STT, and spoken output are split across purpose-specific model components.
Execution	Huenit scripts + robot arm	SVG conversion, drawing, calligraphy, calibration, and pyrography are executed against the physical arm.
Output	OBS, order ledger, shipped artwork	Sessions are recorded, logged, linked back to the customer journey, and turned into video proof and physical delivery.

4. Diagram

Shopify / Telegram / Voice
            |
            v
  chat_ross.py / bob_ross.py
            |
            v
 Claude Haiku / Apertus / Whisper / TTS
            |
            v
 huenit_write / huenit_draw / huenit_svg
            |
            v
      Huenit robot arm
            |
            v
   OBS / order ledger / video proof

5. ATF Overlay

The ATF is layered over the live RobotRoss system rather than replacing it:

• the compiled wiki explains the intended architecture and subsystem behavior
• the operational ledger captures what actually happened in production logs
• the local query path lets an operator inspect the system on the local machine
• the voice interface sits on top of the same evidence-backed query layer

6. Four-Layer ATF Surface

Layer	Component	Description
Layer 1	Compiled Wiki	Structured knowledge pages distilled from docs, code, and architecture notes.
Layer 2	Operational Ledger	Normalized production evidence built from raw RobotRoss log streams.
Layer 3	Local Q&A	A local model interface for answering questions over the wiki and ledger without cloud dependency.
Layer 4	Voice Interface	Optional speech loop using Whisper for STT and Voxtral for spoken responses.

7. Key Runtime Components

• Voice showcase: listen.py performs Whisper STT with VAD, chat_ross.py handles the conversation loop, and spoken confirmations can trigger drawing in the background.
• Main orchestrator: bob_ross.py performs readiness checks, locking, generation, narration, draw execution, and cleanup.
• Drawing pipeline: huenit_write.py, huenit_draw.py, huenit_svg.py, and sketch composition utilities convert requests into robot motion.
• Audio/TTS: the architecture sources mention Kokoro as the primary local neural TTS path, Voxtral for higher-end spoken output, and system-level fallback voices.
• Studio and proof: OBS captures the run, while the order ledger tracks received time, buyer, content, status, and produced video links.

8. Wall of Fame Operating Model

• The commercial target is a 10×10 Wall of Fame with 100 slots.
• A customer buys a slot, submits a prompt or design, the system dispatches the job, RobotRoss draws it live, proof is recorded, and the physical artwork is shipped.
• This commercial journey matters for the ATF because the technical file must explain not only the robot motion, but also the customer-facing intake and proof chain.

9. Notes and Open Points

• The two source architecture documents describe the same system but emphasize different layers: one is voice-showcase and orchestrator centric, the other is end-to-end commerce/studio centric. The ATF page intentionally merges both views.
• TTS references are not fully aligned across the documents: one source emphasizes Kokoro plus system fallbacks, while another also calls out Voxtral. The ATF should present these as runtime variants rather than contradictory claims.
• Some model assignments are mode-specific. Claude Haiku 4.5 appears in control and commerce flows, while Apertus 8B is the local narration/default local reasoning path.
• The compiled wiki still depends on the quality and coverage of the ingested source corpus, so this page should keep being refreshed when the RobotRoss architecture docs change.