Case Study 02 · On-Demand Moving & Urban Logistics Tech

Moovez: Multimodal Vision Quotation & Logistics Engine

Automated Relocation Estimations via Gemini 2.5 Flash, 80+ Item Classifications, Additive Elevator Physics & Streamlit Quotetron

Client
Moovez Canada (Calgary, AB)
Leadership Role
Software Developer Intern (AI & Computer Vision)
Timeline
Jan 2025 — Mar 2026
Primary Tech
Gemini 2.5 Flash
Quantified Executive Impact & Verified Metrics
80+ Items
Catalog Detection
Across 24+ distinct household categories
2–3s
Processing Latency
Per image via single-call Google Files API batch
82%
Volume Safety Buffer
Prevents under-sizing vehicles for irregular freight
62%
Modular Refactor
Codebase reduction (1,188 → 450 lines) with modular OOP
01 / Situation (The Enterprise Bottleneck)

Operating Context & Friction Points

Residential moving quotes historically required customers to manually catalog every household item or wait hours for an on-site dispatcher walkthrough. Inconsistent human estimation caused widespread pricing disputes, truck undersizing, and margin leakage. Furthermore, legacy estimation tools used crude percentage multipliers that over-inflated elevator jobs by up to 68% and failed to distinguish rapid box loading from complex appliance positioning.

02 / Task (The Mandate)

Architectural Scope & Objectives

Engineer an end-to-end computer vision and logistics estimation engine: process mixed media (images, video walkthroughs, iPhone HEIC) with zero client-side degradation, map detected furniture items against proprietary logistics databases (moving_items_logistics_v2.json), calibrate real-world moving friction (replacing arbitrary percentage inflation with additive elevator physics and explicit load/unload category splits), provide interactive crew sliders (2–6 movers) with instant price recalculation, and deploy a production-grade Streamlit application featuring statistical repeatability testing.

03 / System Architecture Topology
Customer Photos & Walkthrough Videos → Google Files API Batch Upload → Gemini 2.5 Flash Vision → 80+ Item Detection (JSON DB) → Additive Logistics Engine (82% Buffer + Elevator Physics) → Interactive Crew Slider & Dynamic Pricing ($min–$max) → Streamlit Quotetron UI.
04 / Architectural Execution (Actions Taken)

Multimodal Ingestion, Batch Files API & Format Processing

Engineered mixed-media batch processing supporting videos (.mp4, .mov, .webm, .avi) and images (including native iPhone .heic/.heif) directly into Google Gemini 2.5 Flash.

  • Leveraged Google Files API to execute a single-call batch upload for all mixed media, slashing processing time by 40–50% compared to per-file requests.
  • Disabled model thinking budget to maximize deterministic JSON inference speed while retaining high-precision item boundary recognition.
  • Detected and categorized 80+ item types across 24+ distinct furniture and household categories with small/medium/large volumetric bounding.

Domain Logistics Physics: Additive Elevator & Explicit Load/Unload Split

Overhauled legacy moving formulas with physical ground-truth logistics rules to ensure stable, defensible estimates across diverse building types.

  • Replaced flawed total-time percentage multipliers with an additive per-trip elevator model (fixed wait/ride times per load, parallelism capped at 1.5 crews), eliminating job inflation.
  • Engineered explicit load/unload time calculations with category ratios: boxes (0.90x), fragile artwork/mirrors (0.80x), heavy appliances/pianos (0.55x), and disassembly items (0.50x).
  • Incorporated stairs multiplicative friction scaling physical fatigue per flight, alongside travel time and regional wage schedules ($170/hr baseline for 2-person crews).

Vehicle Selection, 82% Safety Buffer & Interactive Crew Sliders

Engineered operational vehicle assignment heuristics paired with real-time UI controls for dispatchers and customers.

  • Formulated the 82% Volume Buffer Rule to ensure safe vehicle selection, preventing truck capacity over-utilization for bulky or awkwardly shaped items.
  • Restricted movers to 2 for small vehicles (Pickup, Cargo Van, 10'–12' Truck) while scaling up for 16'–26' trucks based on cab seating and payload limits.
  • Constructed an interactive post-analysis Crew Override Slider (2–6 movers) in Streamlit that dynamically recalculates duration and $min–$max price ranges without re-calling the AI.

Modular Architecture Evolution (V2 → V9) & Statistical Testing

Architected the system through 9 major evolutionary releases, refactoring a 1,188-line monolithic script into an enterprise object-oriented framework with automated testing.

  • Refactored into modular V9 architecture: MoovEZVisionAnalyzerV7 orchestrator, modules/base.py (abstract base classes), file_handlers.py (polymorphism), calculator.py, and ai_client.py, cutting code size by 62%.
  • Deployed the Quotetron web interface on Streamlit Community Cloud with drag-and-drop mixed-media uploads and JSON/CSV export.
  • Engineered an automated 5-iteration statistical test suite rendering variance graphs to validate AI extraction consistency and pricing stability across repeated runs.
05 / Business Results & Enterprise Certainty

Commercial Speed, High-Accuracy Logistics & Live Cloud Deployment

The Moovez Vision Agent transformed residential moving estimation from an error-prone manual triage into an instantaneous, data-driven computer vision experience.

Live production web application deployed and operating on Streamlit Cloud.
Achieved 2–3s per-image analysis and 15–30s video walkthrough processing via single-call Google Files API batch optimization.
Eliminated pricing disputes and volume under-estimations with the 82% safety buffer and additive elevator calibration.
Delivered a maintainable, modular Python architecture with built-in statistical variance analysis for continuous model monitoring.
Production Architecture Stack
Gemini 2.5 FlashGoogle Files APIStreamlit CloudPython 3.11OpenAI GPT-4 VisionComputer VisionLogistics OptimizationObject-Oriented Architecture
Let's Connect

Interested in this project or want to talk tech?

I'm always excited to discuss engineering architectures, share project learnings, or exchange ideas on computer vision, data pipelines, and modern enterprise systems.