GenAI Assignment Completed

[ak1606]

Hi team,

I've completed all 4 problems of the GenAI assignment. Here's a quick summary of what's included:

Problem 1 – Video-to-Notes Pipeline: Compared three approaches (Cloud SaaS, Hybrid Local+LLM API, and Fully Offline). Recommended the Hybrid approach using Whisper API + Gemini 1.5 Pro for best quality/cost balance, with a full batch pipeline design including error handling and timestamp accuracy mitigations.

Problem 2 – LinkedIn Post Generator: Designed a zero-shot prompt that takes a user persona + topic and generates 3 structurally distinct LinkedIn drafts (Concise Insight, Story-Based, Actionable Listicle) as structured JSON in a single API call. Includes API settings, edge case handling, and a user review/approval flow before publishing.

Problem 3 – DOCX Template Bulk Generator: Proposed a system where GenAI (GPT-4o/Gemini) reads a plain Word document and automatically detects variable fields — no manual placeholder tagging required. Covers the full flow: AI detection → human review → single fill → bulk generation via Excel with validation and error reporting.

Problem 4 – Character Video Series Generator: Designed a 5-step episode pipeline (Script → Storyboard → Visuals → Audio → Video) built on top of a persistent Series Bible. Addressed the hardest challenges: character visual consistency via IP-Adapter face embeddings, 5-minute pacing via word-count control, and cross-episode continuity via a character state tracker.

All solutions are practical and implementable with currently available tools and APIs. Happy to discuss any of the design decisions.