Limited research tasks
The system creates only 2–5 focused research tasks so the process stays controlled.
marvinjb.devLIVE AI SYSTEM · RESEARCH AGENT
Turn one question into a clear, cited research report.
Ask one question. The AI breaks it into smaller research tasks, searches multiple sources at the same time, compares the findings, and returns one report with evidence and citations.
LIVEPUBLIC DEMO
01Research tasksBreaks the question into 2–5 research tasks
02Evidence trackingTracks the evidence behind each claim
03Final reportReturns a structured research report
04Production deployedFastAPI · Docker · Nginx
01 · LIVE DEMO
Ask one question. The system does the research.
Ask one question. The AI breaks it into smaller research tasks, searches multiple sources, and combines the findings into one cited report.
Choose a quick or deeper search. The system breaks the question into focused research tasks, searches multiple sources, and shows progress until the final report is ready.
READY
Your completed, source-grounded report will appear here.
02 · HOW IT WORKS
How the research process works
The AI breaks one question into smaller research tasks, searches multiple sources, keeps track of the evidence, and combines everything into one cited report.
Several research tasks run at the same time, then their results are combined only after the evidence is checked.
QUESTIONQuestion
→PLANBreak into research tasks
→SEARCHSearch multiple sources
→EVIDENCECollect evidence
→COMBINECombine findings
→WRITEWrite final report
→REPORTCited research report
03 · ENGINEERING DECISIONS
How the system keeps research reliable
The AI helps research and write the report, but the application controls the evidence, validates the results, and keeps the final report tied to real sources.
Evidence cannot be changed
The AI selects evidence that the application already collected instead of inventing its own evidence.
Results are combined by the application
The system combines the research results using fixed rules instead of letting the AI freely merge everything.
Final report uses validated evidence
The AI writes the report using only approved claims, evidence, and uncertainties.
04 · RELIABILITY
What happens when research fails
The system does not hide errors. If a search or research task fails, successful work can still continue when it is safe to do so, and the final result shows what failed.
Invalid evidence is rejected
Claims or evidence that do not match the collected research are not accepted.
Successful research can still continue
If one research task fails, the others can still contribute to the final report.
Usage is limited
The system limits workers, searches, and requests so cost and load stay controlled.
Errors stay clear
Timeouts, provider failures, bad outputs, and complete research failure are reported separately instead of being hidden.
05 · TECHNOLOGY STACK
FrontendReact · TypeScript
BackendPython · FastAPI · Pydantic · asyncio
AI & SearchOpenAI Structured Outputs · Tavily Basic Search
DeploymentDocker · Nginx · Ubuntu VPS · health checks · CORS · rate limiting
INSPECT THE IMPLEMENTATION
See how it was built
Explore the backend code, tests, and architecture behind the Research Agent.