wip ReAct
Deep Research Agent
An autonomous agent that browses, fact-checks, and generates a PDF report on any topic — with live WebSocket thought streaming and per-report cost metrics.
LangChainWebSocketspdf-libFastAPIOpenAI
A “Market Intelligence” agent that handles long-running autonomous research tasks. The user types a topic — for example, “NVIDIA H200 vs H100 benchmarks” — and the agent does the rest.
What It Does
The agent runs a ReAct loop (Reason + Act):
- Searches ArXiv, web sources, and structured data APIs
- Fact-checks sources against each other
- Synthesises findings into a structured PDF report
Live Thought Stream
As the agent works, every internal step is streamed to the UI via WebSockets:
Thought: Searching ArXiv for H200 benchmarks...
Action: fetch_arxiv("NVIDIA H200 inference benchmark 2024")
Thought: Source A and Source B conflict on FLOPS. Fact-checking...
Action: cross_reference_sources(...)
Thought: Confidence sufficient. Generating summary...This turns the waiting time into a demonstration of agentic reasoning.
Metrics on the Card
Every completed report displays Cost per Report (token usage × model price) and estimated Time Saved versus manual research — making the value proposition concrete and measurable.