Fair, Explainable
Fleet Intelligence
Built for SMB fleets, TraceData addresses 7 critical telematics gaps with a people-first philosophy.
Scroll for Architecture
5 Agents + Shared Tool Gateway
TraceData uses 5 autonomous agents and a shared Tool Gateway to deliver fairness, coaching, burnout detection, appeals, contextual enrichment, and integrated safety-welfare response.
Safety Agent
Processes critical events via queue.critical with a 3-level intervention flow: app notification, formal message, and direct fleet-manager call under a sub-5-second target.
> ACTION: emergency_services + fleet_manager_alert
Scoring Agent
Scores trips and drivers (0-100) using XGBoost, then applies AIF360 fairness mitigation and SHAP/LIME explanations for every decision.
Orchestrator Agent
Entry-point router using deterministic and semantic pathways; invokes ingestion validation sidecar and logs every routing decision for accountability.
Support Agent
Unifies appeals and coaching using pgvector semantic retrieval plus LLM guidance, ensuring contestability and consistent fleet-manager decisions.
Sentiment Agent
Tracks emotional trajectory using a rolling event window and escalates burnout risk alerts to fleet managers with recommended interventions.
Tool Gateway: Context Enrichment
Enriches every inbound event with driver history, route context, and environmental data before agent dispatch.
Tool Gateway: Ingestion Validation
Validates all incoming telemetry schema and filters malformed events before they reach the agent layer.
Fairness-First Operations
Core philosophy: fairness first, driver-centric design, and transparent contestable outcomes with auditable governance.
Explainability
and Assurance Layer
TraceData pairs fairness controls with explainability artifacts so every intervention is observable, reviewable, and contestable across operational and compliance workflows.
LangGraph + Priority Queue Execution
Agent flows run through LangGraph while Celery/Redis routes critical, high, medium, and low events with deterministic response paths.
AIF360 Fairness + SHAP/LIME Attribution
Bias mitigation and feature-level explanations are generated for scoring decisions, with recurring fairness audits for drift control.
Governed Decision Stream
hover to reveal
SLA: < 5s
SPD TARGET: < 0.5
QUEUE: queue.critical
SHAP Influence
Direct Appeals
Drivers contest unfair outcomes with semantic retrieval against precedent cases.
Tone-Calibrated Coaching
Support guidance adapts using sentiment context (encouraging, supportive, or directive).
Operator Dashboard
Every intervention stores decision rationale, FM action, and AI reasoning for traceability.
Human-in-the-Loop
Drivers can contest scores through structured appeals while fleet managers override with context and reasoning logs. The operating model is support over surveillance.
Built for transparent, contestable, and accountable fleet operations.
Technical Specifications
LangGraph + Orchestrator Router
Agents execute as graph nodes with deterministic and semantic routing, plus full routing audit trails.
AIF360 + SHAP + LIME
Bias detection and mitigation are integrated into scoring; SHAP/LIME provides local and global feature explainability.
FastAPI + Celery/Redis + PostgreSQL/pgvector
Async APIs, priority queues, ACID event storage, JSONB support, and semantic vector search for appeal consistency.