Podium AI
A LangGraph-powered agentic research and presentation system that conducts multi-source empirical research, isolates verifiable claims with academic citations, and guides presenters through two Human-in-the-Loop (HITL) gates to generate styled 16:9 PowerPoint decks complete with speaker notes.
The Problem
Most 'AI presentation generators' are glorified text formatters that hallucinate ungrounded bullet points and produce generic slide decks detached from real academic literature. Presenters and educators have zero control over fact verification, cannot audit sources, and end up with static decks that lack speaker talking points or professional visual hierarchy.
Approach
Engineered a LangGraph state machine compiling stateful pause checkpoints (`interrupt_before=['review', 'verify']`) to enforce two distinct Human-in-the-Loop gates before committing compute to slide compilation.
Integrated a resilient multi-source research pipeline combining Tavily and DuckDuckGo search fallbacks, passing raw web intelligence into a dedicated factual extractor that isolates structured claims with citation metadata.
Designed an index-based claim prioritization node that returns 0-based integer rankings (`[3, 0, 5, 1]`) rather than rewriting full text, drastically cutting token consumption and eliminating repetition degeneration loops on compact models.
Built a custom `python-pptx` rendering engine (`deck_style.py`) that applies 4 customizable academic color palettes, calculates responsive shape geometry for 16:9 widescreen slides, and embeds native talking points directly into PowerPoint's speaker notes pane.
Crafted an interactive React frontend featuring an Antigravity glassmorphism design, a live 16:9 Cinema Deck simulator, clickable in-slide web references, in-browser slide editing, and instant Markdown export.
Key Features
- Two-stage Human-in-the-Loop (HITL) review gates for agenda refinement and factual claim verification.
- Multi-angle search engine (Tavily + DuckDuckGo fallback) isolating empirical claims with direct URLs.
- Deterministic 16:9 widescreen PowerPoint compiler supporting 4 bespoke academic visual themes.
- Native PowerPoint speaker notes injection (`slide.notes_slide`) and in-browser Presenter Mode.
- Real-time 16:9 Cinema Deck simulator with in-browser slide editing and one-click Markdown export.
Challenges
The most notorious issue was an LLM degeneration loop during claim prioritization: prompting smaller models (like `gpt-oss-20b`) to rewrite 20+ factual claims verbatim caused them to hallucinate a repetitive 'landscape loop' mid-generation. Transitioning from verbatim text output to an integer index ranking array reduced output tokens from 4,000 to 300, eliminating the bug while speeding up synthesis by 85%. On the presentation side, hand-crafting OOXML shape transitions for PowerPoint required meticulous XML namespace manipulation to avoid presentation corruption across Microsoft PowerPoint, Apple Keynote, and Google Slides. Lastly, managing thread checkpoint state across server cold-starts required custom session storage rehydration and background task cleanup to prevent orphaned PPTX files on disk.
Outcome
A production-grade, verifiable academic presentation platform that bridges research and delivery. Presenters can steer lecture agendas, cherry-pick empirical facts, live-preview themes, and download clean widescreen decks with embedded speaker notes—transforming hours of literature review and slide formatting into a reliable 60-second workflow.