Cognitive / Multi-Agent AI System
PentagonAI - Market Analysis Agent
Solo-built market intelligence backend: ingest → extract structured events (LLM) → canonicalize/merge (Qdrant similarity + LLM) → generate ImpactLinks → multi-agent council forecasts. Includes ops/debug endpoints for model routing and workflow visibility.
- Role
- Solo builder · architecture, implementation, and operational tooling
- Period
- 2026.01 - Present
- Context
- Personal Project
Solo-built market intelligence backend: ingest → extract structured events (LLM) → canonicalize/merge (Qdrant similarity + LLM) → generate ImpactLinks → multi-agent council forecasts. Includes ops/debug endpoints for model routing and workflow visibility.
What made the system difficult.
Required an operable pipeline for ingestion→extraction→canonicalization→council execution, with queue separation and endpoints for ops/debugging so long-running LLM workflows remain manageable.
How boundaries and flows answered it.
Built a FastAPI server with domain routers (events/timeline/council/forecast/triggers/search/threads/ops/auth/users) and a Celery worker with stage-separated tasks. Implemented event identity merging, impact link generation, trigger-driven meeting creation, and forecast resolution/scoring endpoints, plus a model-routing debug endpoint for visibility.
The path of responsibility through the system.
Event Pipeline
Ingestion & Extraction
Responsibilities
- — Multi-source collection
- — Entity extraction
- — Temporal parsing
Analysis Engine
Canonicalization & Linking
Responsibilities
- — Embedding-based dedup
- — Impact link generation
- — Timeline visualization
Overview
- Solo-built market intelligence system: ingestion → event extraction → canonicalization/linking → council-style analysis.
Purpose & User Experience
Purpose
- Convert scattered news/time-series into a navigable event graph (search + timeline) and generate forecast artifacts via a council-style workflow.
User Flow (high level)
- Ingest/backfill → extract events → canonicalize/merge → browse timeline/search → run council → resolve/score outputs and review reports.
Key Components
- Celery worker pipelines separated by stage for operational control
- Qdrant similarity search + LLM judgment for event identity/canonicalization
- Event impact linking and multi-agent “council” orchestration
- Ops/debug endpoints (health, routing visibility) for day-2 operations