BEHIND THE CURTAIN
THE JUDICIAL ENGINE
RAG KNOWLEDGE INGESTION
Process
User text and uploaded PDF whitepapers are processed through a chunking pipeline. Documents are split into overlapping segments of 512 tokens, each converted into dense vector embeddings using the all-MiniLM-L6-v2 sentence transformer model.
These embeddings are stored in a ChromaDB vector database with cosine similarity indexing, enabling sub-millisecond nearest-neighbor search across the entire knowledge base.
User Input → PDF Parser → Text Chunker
↓
Sentence Transformer (MiniLM-L6-v2)
↓
384-dim Vector Embeddings
↓
ChromaDB (Cosine Similarity Index)
SCIENTIFIC GROUNDING
Cross-Reference Protocol
The system performs real-time queries against three major scientific databases: arXiv (physics, CS, math), PubMed (biomedical), and Semantic Scholar (cross-domain with citation data).
Results are deduplicated, embedded, and similarity-ranked against the user's pitch. The top-k=5 most relevant papers are extracted and injected into the LLM context as evidence for scientific validation.
arXiv API → XML Parse → Filter
PubMed API → E-Utilities → XML Parse
Semantic Scholar → REST API → JSON
↓
Deduplication + Re-ranking
↓
Top-5 Papers → Context Injection
MARKET STRESS-TESTING
Analysis Methodology
The system constructs market context by querying the vector database with business-oriented queries derived from the user's pitch. It evaluates unit economics, CAC/LTV ratios, and competitive moats.
Market data is sourced from Y-Combinator deal structures, Gartner reports, and CB Insights market analysis. The RAG pipeline retrieves the most contextually relevant market benchmarks to ground the evaluation.
User Pitch → Query Generation
↓
Vector DB Search (market_reports)
↓
CAC/LTV Analysis
Competitive Moat Assessment
Market Size Validation
↓
Top-3 Reports → Context Injection
GROQ API & LPU ACCELERATION
Technical Overview
THE VERDICT leverages Groq's custom LPU (Language Processing Unit) architecture for ultra-low latency inference. Unlike traditional GPU-based inference, Groq's deterministic single-core execution model eliminates memory bottlenecks and delivers consistent sub-second response times.
The primary model deepseek-r1-distill-llama-70b provides deep reasoning capabilities for complex analytical evaluation, while fallback models (llama-3.3-70b-versatile, mixtral-8x7b-32768) ensure system resilience.
System Prompt + Context
↓
Groq API (openai/v1/chat/completions)
↓
LPU Hardware Acceleration
↓
JSON Response (Score + Critique)
MODELS:
Primary: deepseek-r1-distill-llama-70b
Fallback: llama-3.3-70b-versatile
Fallback: mixtral-8x7b-32768