FaroIQ

The problem

Thousands of nonprofits and community organizations across Latin America make critical decisions without data, without analytical capacity, and without access to strategic consulting. An organization with fifteen volunteers and a twelve-thousand-dollar annual budget cannot afford data scientists or strategy teams. The result is that most of them plan based on intuition, miss funding opportunities due to poorly structured proposals, and run projects without impact metrics that would allow them to learn or scale.

The solution

FaroIQ takes a plain-language description of the organization and its context, runs it through a nine-agent sequential AI reasoning pipeline on Azure AI Foundry, and produces in under ninety seconds a complete set of strategic deliverables ready to use. The pipeline chains each agent's context into the next, accumulating internal coherence at every step. An autonomous revision agent scores output quality and can re-run critical stages if the internal score falls below 0.70. When complete, Azure Logic Apps automatically executes actions across Microsoft 365: creates Calendar events, To Do tasks, uploads documents to OneDrive, and sends the summary via Outlook.

Challenges

• Designing a nine-agent pipeline in strict sequence where each agent receives the accumulated output of all preceding agents without exceeding Azure AI Foundry token budget

• Implementing the autonomous revision loop in the Critique agent, which evaluates coherence, feasibility and completeness of the plan before authorizing Logic Apps execution

• Maintaining an open WebSocket connection throughout pipeline execution for real-time per-stage streaming without blocking FastAPI's event loop

• Managing the differences between Microsoft 365 personal and work accounts for Logic Apps integrations, which have distinct permissions and endpoints

• Synchronizing the Three.js globe visual state with the actual pipeline progress without introducing perceptible latency in the stream

Learnings

• Context chaining between agents is more effective than parallel agents when internal output coherence matters more than raw speed

• An agent that can identify its own weaknesses and correct them autonomously provides more operational trust than one that simply produces output without evaluating it

• Three.js with day/night mode and scene changes requires a full remount via React key prop to guarantee clean state and avoid WebGL memory leaks

• Azure Logic Apps is the cleanest M365 integration path when payload flexibility is needed without depending on a proprietary SDK with complex licensing

• The biggest bottleneck in agent pipelines is not model latency but the serialization and deserialization of accumulated context between steps