Implementation notes, architecture, technical decisions and problems solved — recorded as the project goes rather than reconstructed afterwards.
The current system diagram and how the parts fit together. See flow
Structure, technologies, components and the Supporter-side experience.
APIs, services, data flow, authentication and infrastructure.
LiveKit architecture and the communication flow between participants.
Speech-to-text architecture and speaker identification.
Claude integration, context construction, prompts, outputs and evaluation.
Transcript structure, storage, retention and access.
Hosting, deployment, environments, monitoring and operational requirements.
Each entry: date · problem · investigation · solution · result · next step.
Entries are added as real problems are hit and solved — real-time audio quality, speaker separation, context window construction, latency between transcription and assistance.
Transcription produces context; the model reads that context. The AI is never in the path of the audio between the two people, so a model failure degrades assistance rather than the session.
Real-time communication, transcription, storage and Claude integration first. Marketplace, billing and analytics deliberately deferred.
Product-level decisions are recorded separately under Product.
Initial concept.
Human Supporter + AI model defined.
Professional feedback incorporated.
Core MVP architecture established.
Real-time communication and transcription implemented.
Claude integrated into the Supporter workflow.
Prototype → validation.