Hoshi

• The description states that Hoshi is a voice assistant for one home.
• It runs on a single 16 GB Apple Silicon Mac.
• It uses local inference for speech recognition (Whisper), language generation (Gemma-based model), memory, and speech synthesis.
• It supports spoken interaction through a browser and a Home Assistant Voice satellite.
• Online research is optional and includes sources; it is not presented as current information but as model-labeled and configurable.
• A default-deny capability kernel protects actions that change external state.
• Sensitive systems like memory, cloud access, and speaker recognition are flag-gated.
• The system uses a hexagonal ports-and-adapters architecture with lightweight Python sidecars.
• It includes an opt-in GPT-5.6 research path using OpenAI web search, protected by a daily spending limit.

Note

This is a self-reported description of the product. No evidence of actual deployment or usage beyond the author’s own setup is provided.

• The description states that Hoshi is not intended to be a universal commercial assistant.
• It positions itself as a private, local-first system that understands its own home and exposes how it reached an answer.
• It claims to be honest about every cloud call and still useful offline.
• It emphasizes clear boundaries between local processing and deliberate online research.
• The project also highlights the use of AI agents (Codex with GPT-5.6) in development, including adversarial review and collaboration protocols.

Inference The positioning reflects a niche focus on privacy and local-first design, possibly targeting early adopters or developers interested in personal automation systems. However, no evidence suggests this has evolved into a broader commercial strategy or product roadmap beyond the hackathon submission.

• The description states that Hoshi is shaped around one real home.
• It is described as a personal system, not a commercial assistant for multiple users or households.
• There is no mention of any target customer segment beyond the individual developer or household user.
• No evidence of segmentation or targeting of specific industries or business types.

Note

The ICP appears to be limited to individuals or small households who value privacy and local processing. No evidence of a defined customer persona or market analysis.

• There is no evidence of any pricing structure, monetization model, or revenue streams.
• The description does not mention subscriptions, licensing, or any commercial offering.
• It is described as a personal system built during a hackathon and not intended for commercial use.

Note

No business model or pricing information is provided in the self-reported description.

• Hoshi uses a hexagonal ports-and-adapters architecture.
• Models and infrastructure live behind replaceable interfaces.
• Speech recognition, speaker embeddings, language generation, knowledge retrieval, and speech synthesis are handled by lightweight Python sidecars.
• It integrates with Home Assistant Voice satellite.
• The system includes an opt-in GPT-5.6 research path using OpenAI web search.
• A collaboration protocol called CollabOS is used for AI agent coordination.
• The project was built during a hackathon and includes setup instructions and reviewer verification paths.

Inference Technical architecture suggests a modular, inspectable system with strong emphasis on local processing and safety. However, no evidence of scalability or production-grade delivery beyond the developer’s own environment.

• The project was built during a hackathon (OpenAI 2026).
• It includes setup instructions and a reviewer-verification path.
• No evidence of real-world usage, customer feedback, or adoption.
• The author mentions that Hoshi 0.8 began before the hackathon, but this submission focuses on extensions built during the competition window.

Note

There is no evidence of traction, revenue, or user engagement beyond the developer’s own system.

• No mention of competitors or competitive landscape.
• The description does not reference existing voice assistant platforms (e.g., Alexa, Google Assistant).
• It positions itself as different due to local-first design and privacy features.

Note

No evidence of awareness of or engagement with the broader market for voice assistants or smart home automation.

• The system is described as running on a single 16 GB Apple Silicon Mac, which may limit scalability or usability.
• Speaker recognition was disabled due to safety concerns and lack of reliable offline performance.
• The project is described as a hackathon submission with no commercial intent.
• No evidence of any real-world testing or user feedback beyond the developer’s own experience.

Inference Risk of limited adoption due to narrow use case, hardware constraints, and lack of commercial viability. The focus on local-first design may not translate into mainstream appeal without significant development or repositioning.

1. What is the actual scope of your intended user base beyond the developer’s own home?
2. Are there any plans to scale beyond a single device or household?
3. How do you plan to handle edge cases like speaker recognition or memory retrieval in real-world settings?
4. Is there any intention to monetize this system, and if so, how?
5. What are the limitations of the current architecture that would prevent broader deployment?