Executive Summary
From blind spots to live operational intelligence
The client's drive-through fleet was running on manual timing, fragmented camera/POS systems, and reports assembled after the peak had passed. I led the architecture, build, governance, and rollout program for an edge-first AI control platform that turned every lane into a measurable, real-time pipeline — giving store managers actionable coaching and executives a live view of fleet health, capacity, and revenue opportunity.
The Business Problem
Across 150 stores, every location measured drive-through performance differently — if it measured it at all. Cameras, POS events, audio, and operational reports lived in isolated systems, while regional leaders relied on manual spreadsheets and customer complaints to identify bottlenecks. By the time a slow lane became visible in a weekly report, the peak-hour revenue opportunity had already been lost.
A fleet baseline quantified the structural gap: average lane time was 6:30 against an industry benchmark of 4:15, orders averaged 185 per store per day, manual entry and mishears drove order accuracy down to 87%, and the modeled annual revenue left on the table was $28.5M. Leadership needed more than a dashboard — it needed an operating system that could sense what was happening at lane level, explain why, and trigger action before the customer experience degraded.
The Solution
I designed and delivered an edge-first, cloud-intelligent control platform that combines computer vision, speech recognition, POS events, goal tracking, and predictive alerting. Each store receives a local AI node that handles latency-sensitive vision and audio processing in under one second, while a central cloud control plane aggregates fleet metrics, manages the model lifecycle, and gives regional leaders a live operational view across all 150 locations.
The platform was deliberately designed around action rather than observation: when a lane drifts above target, the system identifies which phase caused the delay and sends a coaching prompt with a recommended next move. When a pattern repeats, it escalates. This shifted the business from reactive reporting to proactive operations without requiring store managers to become data analysts.
How It Works
Sense
Computer vision, speech AI, cameras, and POS events capture every vehicle journey and lane transition in real time.
Understand
Edge inference measures menu, cashier, pickup, and total lane time against store- and fleet-level goals.
Act
Predictive alerts flag drift, recommend coaching, and escalate repeated patterns before they become customer complaints.
Architecture & Governance
The architecture separates latency-sensitive decisions at the store from fleet-wide intelligence in the cloud, while enforcing privacy, role-based access, model governance, and auditability across the full lifecycle.
System Architecture — Store-to-Fleet Control Flow
Five stages · edge-first processing · governed cloud control
Six-camera zones, POS events, and order audio capture the full vehicle journey
Vision, audio, anonymization, and lane-state decisions in under one second at each store
Live metrics and features aggregated across stores for history, comparison, and model learning
Targets, drift detection, coaching prompts, canary rollout, rollback, and role-based controls
Live store KPIs, lane drill-downs, executive summaries, and recommended operational actions
Security, privacy & audit rail runs beneath every stage — on-device redaction, encrypted transport, RBAC, immutable audit logs, retention controls, and policy-driven model updates.
- Edge AI node using local computer vision, audio understanding, anonymization, and offline-first synchronization for latency-sensitive store operations.
- Streaming and feature layer aggregating live metrics across stores for fleet comparison, history, and continuous model improvement.
- Cloud control plane managing goals, alert rules, model registry, canary rollouts, drift monitoring, auto-rollback, and executive reporting.
- Security and governance controls including end-to-end encryption, role-based access, immutable audit trails, configurable retention, and policy-driven updates.
Business Impact & ROI
- Reduced average lane time from 6:30 to 3:25 by identifying the exact phase — menu, cashier, or pickup — creating each delay and triggering targeted coaching.
- Improved order accuracy from 87% to 99.2% through AI-assisted audio/order understanding with staff confirmation rather than unchecked automation.
- Modeled $22.3M in additional annual throughput revenue and $3.4M in dynamic-pricing upside across the 150-store fleet.
- Delivered a fully costed ROI model presented to the executive team: approximately $1.1M deployment investment against $6.16M net first-year benefit after costs, with a 5.6x return and a modeled six-day payback at fleet scale.
- Created live executive visibility across every store, lane, goal, and hour, replacing reports assembled days after the rush with decisions made while the opportunity was still recoverable.
- Established an edge-first rollout model with explicit continue, adjust, or pause checkpoints, allowing leadership to prove value in five pilot stores before accelerating to the full fleet.
My Role
I owned this engagement end-to-end as the client's AI/ML leader, operating at CTO scope: I built the executive business case and ROI model, defined the store-to-cloud operating model, aligned operations, IT, security, and data stakeholders on the control-plane requirements, and set the governance standards for privacy, model lifecycle, rollback, and role-based access.
I also directed the phased rollout logic — foundation, pilot, scaling proof, and accelerated fleet deployment — so leadership had a clear decision gate at each stage rather than a single high-risk "big bang" launch. The result was not just a computer-vision deployment, but an enterprise operating system for drive-through performance that store managers could act on and executives could defend.