Intelligent insights for dealership ops: data layer & forecasting prototypes - ON-1142

Project type: Innovation
Desired discipline(s): Engineering - computer / electrical, Engineering, Computer science, Mathematical Sciences, Mathematics
Company: Leadbox
Project Length: 4 to 6 months
Preferred start date: 01/01/2026
Language requirement: English
Location(s): Ottawa, ON, Canada
No. of positions: 1
Desired education level: Master'sPhDPostdoctoral fellow
Open to applicants registered at an institution outside of Canada: No

About the company: 

Leadbox Inc. is a Canadian automotive marketing Technology company that builds OEM-compliant dealership websites, digital advertising, SEO, and AI-powered merchandising tools. We serve franchise and independent dealers across Canada and in the US, helping them improve inventory turn, marketing efficiency, and customer experience through data and automation. Our platform integrates website analytics, CRM, DMS, call tracking, and inventory data to surface operational insights and power workflows.

Describe the project.: 

Goal: integrate AI into existing dealer infrastructure so that models reliably drive decisions and actions—not demos. We will (1) design an AI integration layer that sits over 6+ data sources (IMS, DMS, CRM, GA4, call logs, market data) with <1-hour latency; (2) deliver production-ready patterns for LLM/RAG and time-series models (inventory risk, response-time gaps, demand signals); and (3) implement secure, observable pathways from “insight” to “action” (dashboards, notifications, and API/webhook triggers).

Intern(s) will build data contracts and embeddings stores for retrieval, define inference APIs (FastAPI), and implement evaluation/monitoring (latency, quality, drift, guardrails). Work includes: schema harmonization, feature pipelines, prompt/knowledge base design, tool-use/agent patterns for constrained tasks, and policy enforcement (PII redaction, access control, logging). We’ll prototype two end-to-end use cases:
• Aging/price risk: model identifies vehicles at risk (age, demand elasticity), recommends action, and emits a task to existing dealer systems.
• Speed-to-lead: model detects SLA breaches, summarizes conversation context, and routes next best action (call/email/script).

Outcomes: (a) a documented AI integration layer (APIs, data contracts, infra diagrams); (b) reproducible evaluation reports (MAPE for forecasts; human-in-the-loop quality for LLM outputs); and (c) 2–3 dealer-facing “insight→action” prototypes wired into current tools, ready for productization.

Required expertise/skills: 

Must-have: Python (pandas), SQL, API design (FastAPI), data modeling, version control, and experience joining messy transactional/event data. Practical ML (time-series, gradient boosting or equivalent) and LLM integration (prompting, retrieval, chunking/embedding, evaluation). MLOps basics: environments, dependency management, CI for data/model code, telemetry/monitoring.

Assets: dbt-style transformations; vector databases; guardrails/safety (PII redaction, rate limiting); experiment tracking; basic front-end (Streamlit) for internal tools; GA4/BigQuery or CRM schemas; experience with observability (OpenTelemetry) and model quality measurement (hallucination/faithfulness tests).