AWS Builder Center

GrizKit — Mobile-First Equipment Intelligence Platform

> A serverless Progressive Web App that transforms how Amazon RME technicians locate parts, diagnose equipment, and reduce mean-time-to-repair (MTTR) in large-scale fulfillment centers — built entirely on AWS for under $1/month.

Sr Mech & Robotics Technician



The Problem

In large-scale Amazon fulfillment and sortation facilities, Reliability Maintenance Engineering (RME) technicians face a recurring operational bottleneck: when a conveyor goes down, the clock starts. Every minute of downtime impacts throughput, packages stop moving, SLA windows shrink, and downstream processes cascade.
The single largest contributor to extended MTTR isn't diagnostics, it's parts retrieval. A technician identifies a failed motor, bearing, or sensor, but then must:
  1. Identify the correct replacement part, Equipment may have a vendor alias number (stamped on the nameplate), an internal EAM asset code (encoded in DataMatrix QR labels), or a hierarchical position ID. These are three separate identification systems that don't cross-reference in any existing tool available to floor technicians.
  1. Cross-reference the part number across multiple systems: Each facility maintains a physical parts store (crib) with thousands of unique SKUs across hundreds of bin locations. But getting from "I can see the manufacturer part number on this equipment" to "here's the internal part number and where it lives in the store" requires logging into the Honeywell/Intelligrated portal to pull the manufacturer part number, then logging into APM (a separate system) to cross-reference it to the internal Amazon Part Number (APN) and bin location. These portals are laptop-only, not available on mobile, and each has its own authentication process.
  1. Physically locate the part: Even after completing the cross-reference, a tech must confirm which bin the part lives in, a process that requires navigating 4-5 screens deep in APM and hoping the bin location data is current.
The net result: What should be a 2-minute "grab the part and install it" task regularly becomes a 15-30 minute odyssey, pulling the laptop out of the toolbox, logging into Honeywell/Intelligrated to get the manufacturer number, then logging into APM to cross-reference it to the APN and bin, navigating multiple screens in each system, and often just asking senior techs who memorized bin locations years ago.



The Solution: GrizKit

GrizKit is a serverless, mobile-first Progressive Web App purpose-built for Amazon RME technicians on the floor. A technician pulls out their phone, scans the equipment's DataMatrix QR code (or types a vendor number), and within 200-400ms receives:
  • Equipment Profile: Full identity: position ID, system classification, area, equipment configuration, and criticality tier
  • Parts Manifest: Every replacement part for that specific equipment, intelligently categorized (motors, bearings, belts, sensors, sprockets, rollers, guards, hardware) with:
  • Part number (Amazon Part Number / APN)
  • Human-readable description
  • Exact bin location in the facility storeroom
  • Part photograph (2,365 images at 96% coverage)
  • Work Order History: Recent maintenance events for context on recurring failures
  • Smart Filtering: Automatically suppresses raw materials, bulk consumables, and obsolete/superseded parts that clutter traditional BOM views














What Makes It Different

Traditional Workflow
GrizKit Workflow
Pull laptop from toolbox, boot up
Pull out phone (already in pocket)
Log into Intelligrated → find mfr. part number
Scan QR or type 5-digit alias
Log into APM → cross-reference to APN + bin
200-400ms — profile, parts, and bin locations returned instantly
Unfiltered BOM (200+ items)
Smart-filtered, categorized (avg. 12-30 relevant parts)
No images — part numbers only
Visual catalog with photos
Manually hunt for bin location in APM
Bin location on every part
Laptop only, not available on mobile
PWA with offline shell, installable on any phone
Total: 15-30 minutes
Total: 15-45 seconds



Architecture

GrizKit is a fully serverless application running in AWS  us-east-1 . The architecture separates concerns into three layers: Delivery, Compute, and Data.

High-Level Flow

Tech's Phone (PWA)
│ HTTPS
▼
Amazon CloudFront ──► Amazon S3 (static assets + part images)
│
▼
Amazon API Gateway ──► AWS Lambda (Python 3.12) ──► Amazon DynamoDB
The Lambda function implements a multi-strategy resolution chain. The same input can arrive as a vendor alias, a QR-encoded asset code, a full hierarchical position ID, or a partial prefix. The resolver cascades through each strategy in order of probability until a match is found, then hydrates the full equipment profile, parts manifest, and work order history from a single-table DynamoDB schema optimized for these access patterns.
All 231,500+ records (equipment profiles, parts associations, lookup indices, work orders, analytics events) live in one DynamoDB table using composite keys and 2 GSIs — no relational layer, no joins, sub-10ms reads.

AWS Services Used

Service
Role
Configuration
Amazon CloudFront
CDN + HTTPS termination
Single distribution, S3 + API Gateway origins
Amazon S3
Static hosting + image CDN
Single bucket, /parts/ prefix for WebP images
Amazon API Gateway
REST API routing
3 routes, CORS enabled, no auth on /lookup
AWS Lambda
Business logic
Python 3.12, 256 MB, 10s timeout, single function
Amazon DynamoDB
Primary datastore
Single table design, on-demand capacity, 2 GSIs

Why This Architecture

  • Cold start ≤ 400ms: Python 3.12 + boto3 (pre-bundled in Lambda runtime) means no deployment package bloat. The function reuses DynamoDB connections across warm invocations.
  • Zero idle cost: On-demand DynamoDB + pay-per-invocation Lambda = the entire stack costs ~$0.50/month at current usage patterns.
  • Global edge delivery: CloudFront serves the PWA shell and 2,365 part images from edge cache. After first load, the Service Worker (network-first with cache fallback) ensures sub-second rendering even on spotty warehouse Wi-Fi.
  • Single-table DynamoDB design: All access patterns (alias lookup, EAM lookup, direct fetch, parts-by-equipment, work-order history) are served by a single table with composite keys. No joins, no relational overhead.



Data Pipeline

GrizKit's intelligence comes from its automated data ingestion pipeline, a multi-source ETL system that unifies disparate equipment databases into a single, queryable knowledge base.

Source Systems

Source
Data Type
Records
APM (Asset Performance Management)
Equipment master, positions, store parts, work orders
3,261 assets, 4,064 positions, 2,463 store SKUs
Intelligrated/Honeywell Portal
Equipment-to-parts mappings (vendor BOMs)
181,785+ part associations
Cross-Reference Engine
Unified alias ↔ position ↔ EAM mappings
42,811 matched pairs
Part Replacement Registry
Superseded/obsolete part mappings
214 replacement records

The Scraping Pipeline

The facility's vendor portal (Intelligrated/Honeywell) contains the authoritative Bill of Materials for every conveyor position, but exposes no API. The Doom Scraper v2 — named after the DOOM Slayer because it rips and tears through vendor data. It's a Playwright-based headless browser automation tool that:
  1. Authenticates against the vendor portal and maintains session state
  1. Iterates through batch files of equipment aliases (processed in configurable batches)
  1. Navigates the portal's ASP.NET WebForms interface (postbacks, viewstate, dynamic grids)
  1. Extracts structured part data: part number, quantity, description, criticality flag
  1. Handles session timeouts with a 3-layer keepalive system (dialog auto-dismiss, background mouse movement, HTML modal detection)
  1. Checkpoints progress for resumability. Can resume from any position after interruption
  1. Outputs structured CSV for downstream processing

Cross-Reference Engine

Raw scrape data is processed through a cross-reference and deduplication engine that:
  • Joins vendor alias numbers to internal position IDs and EAM codes
  • Deduplicates across 5 layers (raw CSV, alias-part pairs, within-batch, against existing DB records, against all historical import manifests)
  • Validates part numbers against the facility's physical store inventory
  • Generates import-ready structured data for batch ingestion into the datastore

Image Pipeline (Argent Eye)

Part photographs are scraped from the asset management portal using a DevTools Console-based extraction tool:
  1. Capture — 2,418 images extracted (1,957 JPG + 461 PNG)
  1. Match — 2,365 matched to facility store inventory (96% coverage)
  1. Convert — Batch PNG→WebP conversion for CDN optimization
  1. Deploy — Upload to S3 under  /parts/  prefix, served via CloudFront



Frontend: Progressive Web App

The client is a mobile-first PWA designed for one-handed operation in an Amazon fulfillment center environment (gloves, poor lighting, urgent pace).

Key Technical Features

  • DataMatrix QR Scanning: Native camera access via html5-qrcode library; supports the DataMatrix format used on facility equipment labels
  • OCR Fallback: For damaged/unreadable labels, optical character recognition extracts numbers from camera feed
  • Intelligent Input Resolution: Accepts any identifier format: vendor alias (5-digit), EAM code (10-digit), full position ID, or partial match
  • Smart Category Grouping: Parts are automatically categorized by the Lambda function's keyword-based classifier into: motors, bearings, belts, chains, sprockets, rollers, sensors, guards, hardware, and raw materials
  • Visual Part Catalog: Each part displays its photograph (WebP, lazy-loaded from CDN) with lightbox zoom
  • Skeleton Loading: Perceived performance optimization: structured placeholder renders instantly while API responds
  • Offline Shell: Service Worker (v14) caches the app shell for instant launch even without network. API calls use network-first strategy with graceful degradation
  • Theme Support: Dark mode (default, optimized for warehouse lighting) and light mode with persistent preference via localStorage
  • Silent Analytics: Non-blocking event tracking via  navigator.sendBeacon()  with batched payloads (5-event threshold or 30-second flush). Tracks: search events, scan events, device fingerprints, position popularity — without impacting UX performance
  • Installable: Full PWA manifest with splash screen, home screen icon, and standalone display mode. Launches like a native app.

Admin Analytics Portal

A companion dashboard (API-key protected) provides operational intelligence:
  • Search Heatmaps: Which equipment positions are searched most frequently (identifies chronic failure points)
  • Usage Patterns: Peak activity times, unique device counts, scan vs. type ratios
  • Coverage Metrics: Track which facility areas have full data coverage vs. gaps








Operational Impact

Quantified Time Savings

Metric
Before GrizKit
After GrizKit
Improvement
Part identification
5-15 min
15-30 sec
~95% reduction
Bin location lookup
3-8 min
Instant (included in results)
~99% reduction
Full repair cycle (part retrieval)
15-30 min
1-3 min
~90% reduction
New tech onboarding (parts knowledge)
Weeks of shadowing
Self-service from day 1
Eliminated
Cross-referencing part numbers
Manual (senior tech knowledge)
Automated (42,811 mappings)
Eliminated

Scale

  • 231,500+ DynamoDB items — Equipment profiles, parts associations, work orders, lookup indices, and analytics events
  • 2,365 part photographs — 96% coverage of the facility's physical store inventory
  • 42,811 cross-referenced part mappings — Vendor BOMs unified with internal inventory
  • 4,064 equipment positions — Full facility coverage across all sortation areas
  • ~$0.50/month operational cost — True serverless economics at facility scale



Design Principles

  1. Phone-first, gloves-on: Large touch targets, high contrast, minimal navigation depth. A technician at a downed conveyor should reach their answer in ≤2 taps.
  1. Single source of truth: Multiple identification systems (vendor alias, EAM, position ID) all resolve to the same equipment record through the Lambda resolution chain. No ambiguity.
  1. Intelligent filtering: The Lambda function's keyword classifier automatically suppresses noise (raw materials, bulk hardware, paint, adhesives, sheet metal) and surfaces only actionable replacement parts — reducing average BOM views from 200+ items to 12-30 relevant parts.
  1. Infrastructure as cattle: Single-table DynamoDB, single Lambda function, single S3 bucket. No clusters, no servers, no patching. The entire stack can be redeployed from a single shell script.
  1. Progressive enhancement: Works as a basic web page on any browser. Enhanced with PWA capabilities (offline, install, camera) when supported. Degrades gracefully on legacy devices.



Technology Stack

Layer
Technology
Rationale
Frontend
Vanilla JS + HTML5 + CSS3
Zero build step, no framework overhead, instant deployability
Scanning
html5-qrcode (DataMatrix)
Native camera API, no native app required
Caching
Service Worker (v14)
Offline-first shell, network-first API strategy
CDN
Amazon CloudFront
Global edge, HTTPS, WebP image optimization
Storage
Amazon S3
Static hosting, image CDN, zero-ops
API
Amazon API Gateway (REST)
Route multiplexing, CORS, throttling
Compute
AWS Lambda (Python 3.12)
Cold start <400ms, zero idle cost
Database
Amazon DynamoDB (single-table)
Single-digit ms reads, on-demand scaling, no ops
Scraping
Playwright (Python)
Headless browser automation for vendor portal ETL
Image Processing
Pillow (Python)
Batch PNG→WebP conversion
Analytics
sendBeacon + DynamoDB
Non-blocking, batched, zero-overhead telemetry
IaC
AWS CLI + shell scripts
Lightweight, auditable, repeatable



Security Model

  • No authentication on /lookup: Deliberate design choice. The data (equipment BOM, bin locations) has no sensitivity classification. Removing auth eliminates friction for technicians who share devices and work in shifts. The endpoint is rate-limited at API Gateway.
  • API key on /admin/stats: Analytics dashboard requires an admin API key to prevent data exfiltration of usage patterns.
  • HTTPS everywhere: CloudFront enforces TLS 1.3; no plaintext communication path exists.
  • No PII stored: Device IDs are anonymous fingerprints (localStorage UUID). No user accounts, no names, no login sessions.
  • Principle of least privilege: Lambda execution role is scoped to a single DynamoDB table (read for lookups, write for analytics ingest only). No access to source systems, no cross-account permissions.



Cost Analysis

Resource
Monthly Cost
Notes
DynamoDB (on-demand)
~$0.25
Read-heavy workload, items cached aggressively
Lambda
~$0.10
~5,000 invocations/month at 256MB/200ms avg
API Gateway
~$0.02
$3.50/million requests
S3
~$0.05
~500MB storage + GET requests for images
CloudFront
~$0.08
Data transfer out (most hits served from edge)
Total
~$0.50/month
Full-facility intelligence platform
For context: A single minute of conveyor downtime at a high-volume sortation facility costs orders of magnitude more than GrizKit's annual operating budget.



Roadmap

Completed (Phases 1-6)

  • ✅ Core equipment lookup (alias, EAM, position ID, prefix)
  • ✅ Parts manifest with bin locations and smart categorization
  • ✅ Work order history integration (in progress)
  • ✅ Visual parts catalog (2,365 images, 96% coverage)
  • ✅ DataMatrix QR scanning + OCR fallback (OCR currently in floor-testing phase)
  • ✅ Silent analytics + admin dashboard
  • ✅ PWA (offline shell, installable, Service Worker v14)
  • ✅ Full-facility data coverage (~1,895 positions scraped, 231K+ records)

In Development

  • 🔨 Interactive Facility Map ("Wayfinder"): SVG-based spatial interface showing equipment locations, conveyor paths, and store bin positions overlaid on the facility floor plan
  • 🔨 Search Heatmap: Visualization of which equipment positions are searched most frequently, surfacing chronic failure patterns for predictive maintenance planning
  • 🔨 Building-wide Full Sweep: Completing the final ~939 positions for 100% BOM coverage

Future Vision

  • 🔮 Predictive Failure Correlation: Analyze work order + search frequency patterns to predict imminent failures before they cause downtime
  • 🔮 Multi-facility, Vendor-Agnostic Expansion (in progress) — Templatize the architecture for deployment across multiple Amazon fulfillment and sortation sites regardless of equipment installer — Intelligrated, Dematic, or any other vendor. The data model and resolution chain are designed to be vendor-neutral; only the scraping pipeline adapts per vendor portal.
  • 🔮 Motor Fault Telemetry: Integration with vibration/current monitoring data for condition-based maintenance triggers



Builder Takeaways

  1. Start with the user's hands: Literally. GrizKit was designed around a technician holding a phone in one hand and a wrench in the other. Every architectural decision traces back to: "Does this make the tech faster?"
  1. Single-table DynamoDB is underrated: 231K+ items, 5 distinct access patterns, zero joins, sub-10ms reads. A well-designed composite key schema eliminates the need for a relational layer entirely.
  1. Serverless at scale is cents, not dollars: A full-facility intelligence platform serving thousands of lookups for $0.50/month proves that the "serverless is expensive at scale" myth doesn't hold for read-heavy workloads with proper caching.
  1. PWA > Native for industrial use: No app store approval, no device provisioning, no MDM policies. A URL works on every phone a technician owns — personal or company-issued.
  1. Automate your data pipeline early: The Doom Scraper paid for itself immediately. Manual data entry of 181,785 part mappings would have taken months. Automation with Playwright completed it in days with higher accuracy.


Built by Jonathon Dancy | Sr. Mechatronic & Robotics Tech | Amazon MEM1 — Memphis, TN Architecture: AWS Lambda + DynamoDB + CloudFront + S3 + API Gateway Runtime cost: ~$0.50/month | Development time: Iterative over 6 phases
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article