Overview
Applied AI · Multimodal agentic systems
AtlasMind: Multimodal Agentic Research Assistant
An agentic research assistant that reasons across text, images, and live web sources to produce structured, cited outputs.
5 tool types · zero hallucinated citations01Multimodal Ingestion
02Planner Agent
03Tool Router
04Synthesis Engine
05Cited Report
My scope
The boundary I owned.
I designed the planner–executor loop, multimodal context assembly, tool-routing layer, and the grounded synthesis pipeline.
Engineering evidence
- Built a LangGraph planner–executor loop that decomposes research queries into typed sub-tasks and routes each to the correct tool: web search, image analysis, document retrieval, or code execution.
- Assembled multimodal context windows by embedding images through CLIP and interleaving visual tokens with retrieved text before each synthesis step.
- Enforced citation grounding by binding every claim in the final report to a source node in the execution graph, eliminating unsupported assertions at output time.
OutcomeEnd-to-end research reports from a single prompt, with traceable source attribution.