Back to portfolio

Internal Multi-Agent Research Platform

Built an internal deep-research platform at Inspire11 to replace ad hoc prospect research with a repeatable, template-driven agentic workflow for company analysis and business-development support.

What I led

Designed and implemented the full agentic research system (planning, search, synthesis, quality gating, and response assembly), partnered on AWS hosting rollout, and handed off ongoing updates to the broader team.

Stack

Agentic workflow orchestrationTemplate decomposition and structured response planningTavily and complementary web/document search APIs12 custom agents for search quality, sufficiency checks, and workflow controlAWS-hosted deployment servicesDynamoDB persistence
+2 more
Temporary vectorized index for retrieved filesduring response buildingFront-end metadata payload generation for report display and summary views

Highlights

  • Built a template-first research flow where each run starts with a decomposition step that structures output goals before any search execution.
  • Implemented multi-agent search orchestration with dedicated agents for web discovery, document-focused retrieval, and intelligent company-site navigation (board, leadership, subsidiaries, and partnerships).
  • Designed manager/editor-style synthesis and revision passes with explicit quality gates before final output generation.
  • Added file handling for retrieved artifacts and a temporary vectorized retrieval layer so source files could be referenced during report assembly.
  • Implemented persistence and traceability patterns in DynamoDB for templates, searches, and result state.
  • Added search retry controls with quality assessment and retry limits to balance reliability with API cost.
  • Produced structured output bundles (template-aligned report plus metadata summary) for front-end consumption.

Outcomes

  • Established a reusable internal agentic research capability adopted across teams.
  • Platform is actively used internally as a major initiative within the company, with regular usage by a majority of the BD team.
  • Improved internal ability to run structured deep-research workflows with reusable templates and agent controls.
  • Platform serves approximately 10 users in the BD/research context.
  • Approximately 90 research reports were generated over the past 3 months.
  • Positive user acceptance and recurring feedback loops indicate the tool is considered essential for project kickoff workflows.
  • Shifted research behavior from ad hoc manual ChatGPT sessions to a shared internal workflow for repeatable prospect evaluation using public information.