Engineering Audio Reviews and Agentic Workflows with Miadi

Source: miadi-screenwalk-2610061207-what-bengio-implied-and-the-mia-ava-podcast — JGWill
https://youtu.be/U4FGtwCsosc
type: internal

Summary

Guillaume demonstrates the architectural expansion of the Miadi review system, evolving it from static document processing into an interactive agentic ecosystem. By integrating transcript handling, local execution scripts, the new 'apply' action for grounded research prompts, and the automated 'mia-ava-podcast' synthesis skill, developer sessions become listenable, verifiable, and strictly human-approved.

Aspects

Local Execution Scripting (10:18-10:40)

To bypass the 60-second execution timeout on hosted services when processing long screen recordings, a local script execution environment was built. It interfaces directly with the shared database, enabling local execution while publishing validated outputs publicly.

Transcript Ingestion & Caching (10:51-11:35)

The client added commands to store, read, and generate transcripts for review sessions. Re-running revisions against saved transcripts avoids costly video re-evaluations, caching ground truth while tracking speaker turns and silent intervals.

Prompt Boundary Separation & Pragmatics (19:40-20:30)

To safely isolate agent commands from review metadata, the system uses syntactic boundary detection. Structured review targets are separated from plain-text prompts, treating LLM output as data rather than executable instructions.

Multi-Disciplinary Grounding (16:20-18:08)

The agent framework incorporates software engineering interface compatibility, computational linguistics normalization, archival provenance tracking, and human-computer interaction principles to ensure trustworthy, structured reviews across iterations.

End-results and Core Aspect

Core Aspect — Human-Supervised Synthesis Loop (17:45-18:08)

The core implementation establishes a loop where the agent parses review data, normalizes terminology, and drafts updates, but publishes nothing until the human reviewer verifies and approves the exact proposed changes in a structured status table.

Results

• Multi-Review 'Apply' Action (13:23-14:15): The newly added 'apply' action inspects several review IDs alongside user prompts, extracting terminology to build foundation packets without unapproved side effects.
• Automated Audio Review Synthesis (22:42-23:50): The 'mia-ava-podcast' skill transforms complex developer logs into validated, multi-turn dialogues, verified by automated dialogue checkers for word counts and structure.