docs/progress.md

Progress

2026-06-08

Completed

• Confirmed Agent Tina is a FastAPI app, not a Streamlit app.
• Added local setup helpers:
  • setup.ps1
  • run.ps1
• Verified local dependencies, ffmpeg, app import, and local Uvicorn startup.
• Switched default Whisper model from tiny to base.
• Added Hugging Face Spaces Docker deployment support:
  • Dockerfile
  • Space metadata in README.md
  • deployment script at scripts/deploy_hf_space.py
• Deployed Space:
  • Repo: https://huggingface.co/spaces/KPrashanth/agent-tina
  • Live app: https://kprashanth-agent-tina.hf.space/
• Confirmed Hugging Face Space status: RUNNING.
• Updated UI for the agreed meeting workflow:
  • meeting code
  • participant/device name label
  • meeting type
  • recording mode
  • known terms for future correction
  • same-room recording warning
  • raw transcript, corrected transcript, and MoM output sections
  • Markdown download buttons
• Updated browser page title from Voice MoM Agent to Agent Tina.
• Implemented the meeting-session MVP:
  • host-created meeting id and private host key
  • participant join links and host dashboard links
  • participant audio segments stored under meetings
  • Personal Mic and Room Recorder modes
  • room label and covered participants
  • LLM transcript correction for meeting segments
  • final MoM generation across all segments
  • server-side raw transcript, corrected transcript, and MoM Markdown files
  • host-only Markdown downloads
  • optional Hugging Face Dataset sync
• Verified the deployed workflow end to end:
  • created a meeting through the live Space API
  • uploaded sample.wav
  • transcribed with Whisper
  • corrected transcript through OpenRouter
  • generated final MoM
  • downloaded minutes.md
  • confirmed metadata.json, raw transcript, corrected transcript, and MoM exist in the Dataset repo

Product Direction Agreed

• Save raw transcripts, corrected transcripts, and MoM as Markdown files.
• Add LLM-based transcript correction before MoM generation.
• Support multi-device collaboration with participant name labels.
• Do not start with automatic speaker detection or voice diarization.
• For in-person meetings, prefer one Room Recorder device instead of every participant recording in the same room.
• Final build plan documented in docs/final-build-plan.md.
• Host-created meeting sessions will be the next major product step.

Next Implementation Target

Polish and harden the deployed workflow:

• improve long-running transcription progress/error states
• add meeting close/archive controls
• show Hugging Face Dataset sync status
• add regression tests for storage and API routes

2026-06-09

Completed

• Evaluated data/recordings/day-end-meet-08-06-2026.aac.
• Confirmed approximately 10-minute stereo AAC room recordings are supported.
• Added context-guided Whisper transcription using meeting context, known terms, room label, and covered participants.
• Increased default Whisper beam size to 5 for improved accuracy.
• Added original audio preservation under each meeting folder.
• Added optional original-audio Dataset sync.
• Added host-only original-audio downloads.
• Hardened host segment rendering against participant-supplied HTML.
• Added transcript correction content-loss detection and raw-text fallback.
• Separated correction and MoM models:
  • correction: openai/gpt-4o-mini
  • MoM: openai/gpt-oss-20b
• Verified correction on the AAC sample with approximately 97.7% word retention in about 25 seconds.
• Removed transcripts, outputs, and host keys from the public meeting endpoint.
• Enforced private visibility for the meeting Dataset repo.
• Increased meeting-id entropy to make shared join links harder to guess.
• Verified the final deployed AAC workflow:
  • approximately 10-minute AAC Room Recorder segment processed in about 105 seconds
  • corrected transcript retained approximately 98% of raw transcript words
  • public meeting endpoint exposed no transcript segments
  • original 9.5 MB AAC downloaded successfully through host-only route
  • original audio, metadata, raw transcript, corrected transcript, and MoM persisted to the private Dataset repo
• final MoM generation completed in about 8 seconds
• Made long, multi-speaker Room Recorder audio the primary workflow.
• Added a configurable 250 MB upload limit.
• Added audio duration, format, channel, and sample-rate metadata.
• Serialized Whisper transcription to protect CPU reliability under concurrent submissions.
• Fixed stale-snapshot merging so simultaneous participant submissions do not overwrite each other.
• Persisted original audio and uploaded status before long transcription begins.
• Made in-person meetings and Room Recorder the UI defaults.
• Required room label and covered participants for Room Recorder submissions.
• Added atomic locked meeting storage for concurrent segment/status updates.
• Added Dataset restore-on-demand after Space rebuilds.
• Added lazy recovery so meeting metadata restores without downloading large audio until requested.
• Added strict meeting-id path validation.
• Verified deployed post-rebuild recovery:
  • older meeting metadata restored from the private Dataset
  • 9.5 MB original AAC restored lazily on host download
  • invalid meeting paths returned 404