Decision Graph AI

Agentic decision-planning workspace with branching reasoning

// overview

An agentic decision-planning workspace that expands a user prompt into multiple reasoning branches and synthesizes the strongest recommendation through LLM prompt chaining and tree-structured planning. A full-stack React/FastAPI/PostgreSQL interface persists decision graphs, branch memory, and semantic context for retrieval-augmented refinement and versioned exploration.

// architecture

A React frontend renders the decision graph and streams updates over WebSockets as branches expand. A FastAPI backend drives the LLM prompt-chaining and tree-structured planning workflows, while PostgreSQL persists decision graphs, per-branch memory, and semantic context so prior reasoning can be retrieved and refined. Versioning lets users revisit and compare earlier explorations.

// challenges

• ·Structuring LLM reasoning as an explorable tree of branches rather than a single linear response, then synthesizing one strong recommendation from it
• ·Persisting branch memory and semantic context in PostgreSQL so retrieval-augmented refinement stays coherent across versions
• ·Streaming incremental branch updates to the React client over WebSockets without overwhelming the interface

// learnings

• ·Making model reasoning branch and persist turns a one-shot answer into something a user can actually interrogate and refine
• ·Versioned decision state is what makes exploration useful — being able to revisit and compare branches changes how people decide
• ·Tree-structured planning needs a clear synthesis step, or branching just produces more options without a recommendation