How shashwat-akhilesh-shukla/pitchpilot works
Project type and purpose
• PitchPilot is an open-source, multi-agent AI toolkit for end-to-end startup pitch deck creation.
• It orchestrates specialized agents (research, competitor analysis, pitch drafting, slide design, visuals), uses semantic memory, and exports PowerPoint presentations.
File structure
• Top-level modules:
– agents/ – Agent implementations and orchestrator
– memory/ – Qdrant-backed semantic memory
– prompts/ – LLM prompt templates per agent
– utils/ – Helper functions (context prioritization, reflection loops, PPTX exporting)
– custom_llm.py – LLM wrapper plugin
– config.py – Global settings (LLM model, memory, template list)
– main.py – Workflow entry point
– app.py – (Optional) Web/CLI front door
– templates/ – PowerPoint templates
README for insights
• Workflow stages: user input → research → competitor analysis → content drafting → reflection loops → slide design → visuals → PPTX export.
• External dependencies: Hugging Face LLMs, Qdrant, optional Perplexity API, Matplotlib, python-pptx.
Main components & interactions
a. User Interface Layer
– CLI or app.py takes startup info (founders, market, goals).
– main.py loads config and invokes Orchestrator.
b. Orchestrator
– Coordinates sequential execution of agents.
– Passes context and memory between agents.
c. Agents
– ResearchAgent: market research + memory queries
– CompetitorAnalysisAgent: side-by-side analysis
– PitchCreationAgent: slide-by-slide content generation
– SlideDesignAgent: layout, titles, and placement
– VisualGenerationAgent: charts (Matplotlib) & business visuals (image APIs)
d. Memory Layer
– QdrantMemory (memory/qdrant_memory.py): store/retrieve vector embeddings for context recall.
e. Prompt Layer
– Each agent uses its own prompts module to format LLM calls.
f. LLM Layer
– custom_llm.py abstracts interaction with Hugging Face models (configurable model & temperature).
g. Utility Layer
– context_prioritization, reflection_loops, memory_pruning, presentation_exporter.
h. Output Layer
– presentation_exporter writes final .pptx using python-pptx + templates/.
Architectural patterns & principles
• Modular Agent-based design: clear separation of concerns per task.
• Pipeline orchestration: linear workflow via Orchestrator.
• Plugin-style extensibility: custom_llm, prompts, templates.
• Semantic memory caching & pruning: decouples knowledge store from agents.
• Reflection loops: self-critique + iterative improvement.
Final checklist
– [ ] Components grouped and labeled
– [ ] Notation of external services & data stores
– [ ] Callouts for extension points