Syncretos
Founder
May 2026 - Present
Designing a Python system where three autonomous LLM agents independently research a question, collect documents and observations, and merge their collections into a shared evidence base for cited answers and follow-up questions.
The evidence model preserves original documents, table values, measurement conditions, and exact source references. It retains conflicting observations, separates source claims from derived conclusions, and deduplicates documents without losing provenance. Follow-up answers can retrieve supporting passages from the original sources.
Compared PydanticAI, LangGraph, and SGR Agent Core for tool use, agent state, and interruption recovery. Research covers lightweight working memory, SQLite/FTS5 retrieval, and DSPy for evaluating extraction accuracy and whether citations support answers.
Evaluated Tavily and OpenAlex for source discovery, Trafilatura and Firecrawl for web extraction, and Docling, LangExtract, and Mistral OCR for preserving text, tables, and source locations during document processing.