Skip to main content
AI / Backend Systems·PROJECT_01

AI Decision Routing Engine

Domain
AI / Backend Systems
Role
Full-Stack AI Engineer
Stack
FastAPI · Python · React + Vite · Groq LLM · ModernBERT
Year
2026
Conduit project cover

Overview

Conduit is a high-performance decision routing system that classifies and triages unstructured payloads in real-time using a two-layer AI stack: a lightweight local ModernBERT router (Laya) for zero-cost on-device inference, and Groq-hosted llama-3.3-70b for complex reasoning.

FastAPIPythonReactViteGroqModernBERTPydantic v2
The Challenge

Support and operations teams drown in unstructured ticket floods. Manual triage is slow, inconsistent, and doesn't scale. The challenge was building a system that routes tickets to the right priority (P0–P3) with confidence scoring and graceful uncertainty escalation.

System Architecture

FastAPI backend with async request handling. Input sanitization via BeautifulSoup4. Dual-model routing: ModernBERT for cheap local classification, Groq LLM for full reasoning. Pydantic v2 enforces strict JSON schema on every response. React + Vite enterprise dashboard for real-time triage visualization.

Validation Metrics

Routing Engine
Dual-Model (ModernBERT + Groq)
LLM Reasoning
llama-3.3-70b-versatile
Response Latency
~780ms API / <15ms Local
Confidence Tiers
4-Level SLA Triage (P0–P3)
Key Highlights
Zero-crash architecture — graceful fallback on API timeout or parse failure.
Confidence < 0.5 automatically escalates to human review queue.
ModernBERT handles privacy-sensitive payloads entirely on-device.