Democratizing data-driven expertise for small business owners.

The modern operational environment for businesses is characterized by immense complexity. Entities must navigate a vast, interlocking web of constraints spanning multiple jurisdictions—federal, state, and local—covering everything from stringent employment law to evolving consumer safety standards and taxation requirements.

Aleithia is designed to bridge the inequity between large firms with massive legal resources and small business owners who need the same data-driven insights to thrive.

Aleithia is an intelligent platform that:

• Aggregates disparate regulatory, political, and consumer sentiment data.
• Analyzes data using advanced AI/ML to identify critical risks.
• Translates complex findings into actionable, context-specific recommendations.

The backend architecture leverages Modal’s asynchronous job functionalities. A swarm of jobs fetches news across relevant sources in Python, scheduled periodically, while distributed AI inference parses this data into detailed assessments in seconds. The deployed API is now organized into modular route and service packages under modal_app/api/.

The backbone of Aleithia consists of 33 serverless functions leveraging 21 distinct Modal features:

• GPU Inference: * Qwen3-8B AWQ (INT4) self-hosted on H100 via vLLM.
  • <3s cold starts via @modal.enter(snap=True) memory snapshots.
  • bart-large-mnli + roberta classifiers on T4 GPUs using @modal.batched (10x cost efficiency).
• Computer Vision: YOLOv8n (CCTV) and SegFormer-b5 + YOLOv8m (Satellite parking) on T4 instances.
• Regulatory Retrieval: Live Legistar and Federal Register lookups are deduplicated against cached volume data, with cache fallback when upstream sources fail.
• Event Bus: modal.Queue decouples CPU-bound ingestion from GPU enrichment.
• Recursive Agents: Lead Analyst scores docs; high-impact events (7+/10) trigger .spawn() fan-out to 4 parallel worker agents.
• Self-Healing: modal.Period reconciler checks health metrics every 5 min.

Replaces the traditional Postgres + Pinecone + Redis stack with a single Graph-RAG system:

• Hierarchical Isolation: Separates chicago_data (world knowledge) from user_{id} (personal context).
• State Mutation: Handles temporal updates (changing regulations) without re-indexing.
• Context Injection: Automatically connects Reddit sentiment to City Council hearings for personalized responses.

Built with React 19 + TypeScript + Vite + Tailwind CSS v4, featuring:

• Intelligence Briefing: Structured neighborhood intelligence with regulatory, market, and community signals.
• Risk Scoring: ISO 31000-aligned, 6 MCDA dimensions with sigmoid-normalization.
• Visual Intelligence: Satellite parking detection and live IDOT highway traffic cards.
• Knowledge Graph: Interactive drag-and-filter node visualization.
• Professional Exports: 9-section "Investment Committee" proposal format in PDF.

• Heatmap Rendering: Required a custom compression pipeline for Mapbox GeoJSON to aggregate per-neighborhood stats.
• Data Scarcity: Solved for lack of street-level CCTV by sourcing "walk-in potential" from CTA L-station ridership data.
• Observability: Built W3C context propagation from scratch to link spans across the serverless orchestrator.

• Global Expansion: Identifying higher-quality data sources beyond the Chicago area to support global regions.
• Predictive Analytics: Developing deeper anomaly detection to provide founders with prospective locale-specific insights.