HardwareScan

AI-powered hardware identification and inventory for ITAD / e-waste operations — in a single HTML file.

Scan a barcode (or photograph a bare, unlabeled part), let an AI provider identify it with live web search, and build an export-ready inventory — grading, lot tracking, data-destruction status, and a local knowledge base that gets smarter every time a lookup fails and a human fills in the gap.

No build step. No server. No install. One file, open it in Chrome, go.

▶ Live demo — click "Or try the demo — no key needed" on first launch. Sample inventory loads instantly and lookups are simulated locally.

---

Why this exists

Component-level ITAD work has an identification problem. Whole machines flow through ERP systems with service tags and serial databases — but the loose parts bin doesn't. A tub of mixed RAM means typing part numbers into a search engine one stick at a time, and the industrial/embedded parts (Advantech, Innodisk, ATP…) often return nothing at all. Bare sticks with no label don't even get that far.

HardwareScan turns that into: scan → AI identifies it with web search → press Enter → next part. And when the AI can't identify something, the manual-entry fallback teaches the tool — the next time anyone scans that barcode, it resolves instantly from local data, forever, for free.

Features

Identification

• USB barcode scanner input (keyboard-wedge) with part-number cleaning and a transparent Scanned: X | Sent to AI: Y trace
• Four AI providers — OpenAI, Anthropic, Gemini, Groq — each using live web search, with per-hardware-type source steering (industrial RAM vendors, Intel ARK, TechPowerUp, FCC ID…)
• Photo identification for unlabeled parts: the vision model reads chip silkscreen codes and web-searches them in the same call
• Auto-capture with two trigger models: handheld (buffers your sharpest frame over a burst window — hand tremor is irrelevant by design) and bench (place part, hand withdraws, hold-still snap). Adaptive crop: the detected object defines its own capture region, so a SODIMM and a 3.5" drive both work untouched

A tool that learns

• 30-day cache for AI lookups — repeat scans are instant and free
• Permanent memory for human-verified data: failed scan → manual entry → that barcode is remembered forever, including label/serial codes that map to a different orderable part number
• Perceptual-hash photo memory: re-photograph a previously identified part and it offers the stored result (confirm-first — visually identical sticks can be different models, so a human always ratifies the match)
• Rapid count: re-scanning a part already in the current lot just bumps quantity — no AI call, no API key needed, no keystroke

Inventory & workflow

• Condition grading A–D (keyboard-driven: A/B/C/D + Enter, with a scanner-burst guard), toggleable per-job
• Lot / customer / location session stamping; data-destruction status (pending / wiped / certified) with bulk updates
• Batch queue mode with per-item grade override and quantity steppers; manual entry form; inline cell editing; undo
• Search, filter chips, sortable columns, collapsible stats (totals by type/grade/lot, total RAM GB, total storage TB)
• Export to Excel/CSV (three templates, by-lot filter); CSV/Excel import with column mapping and preview
• Dark mode (default), large-font mode, sound cues designed to be parsed eyes-off

Quick start

1. Download HardwareScanGit.html (or clone and open it). Chrome recommended.
2. Open it. Click "Or try the demo" to explore with sample data — or:
3. Settings → pick a provider → paste an API key. Groq has a free tier; keys live in your browser's localStorage and are sent only to that provider's official API endpoint, nothing else.
4. Scan tab → scan or type a part number → Enter.

Sample part numbers that work in demo mode: KVR16N11/8, M378A1K43CB2-CRC, ST4000DM004, SR2L6, WS-C2960X-24TS-L.

Architecture

The interesting design constraint: the identification logic is built for a future beyond the browser.

┌──────────────────────────────────────────────┐
│  UI layer — DOM, localStorage, workflows     │
│  (scan tabs, inventory, caching, settings)   │
├──────────────────────────────────────────────┤
│  LOOKUP CORE  — zero DOM dependencies        │
│  lookupHardwarePart()   pure async           │
│  lookupHardwareImage()  pure async           │
│  parsePartNumber · detectHardwareType        │
│  buildLookupPrompt · normalizeFields         │
│  per-provider request builders + parsers     │
└──────────────────────────────────────────────┘

Everything inside the LOOKUP CORE markers takes explicit context (provider, key, model, field definitions) and touches neither the DOM nor storage — caching is a UI-layer concern. The core is designed to be lifted verbatim into a Node.js MCP server, so the same identification engine can serve every workstation from a shared, learning parts database. See docs/ARCHITECTURE.md for the full writeup and extraction guide.

Other deliberate choices documented there: why photo lookups can't cache like barcodes (and the perceptual-hash + human-confirmation design that works instead), the scanner-burst keyboard guard, and the two auto-capture trigger models.

Known gotchas (field-tested)

• OpenAI "quota exceeded" that isn't. Some org/project configurations restrict spend on legacy-tier models and report it as insufficient_quota — a billing-shaped error for an access-shaped problem. Symptom: text and photo calls fail with a quota message while the billing page shows credit. Fix: use a current-generation model (default here is gpt-5.4-mini).
• Anthropic from the browser requires the anthropic-dangerous-direct-browser-access header; server tools (web_search) don't support forced tool_choice.
• Barcode ≠ part number. RAM labels carry multiple barcodes; the scannable one is often a serial or agency code. The tool's alias memory exists precisely for this — identify it once manually, and that label resolves forever.
• Webcam optics are the floor for photo ID. Chip silkscreen is ~1mm text; fixed-focus laptop cameras have a sweet-spot distance. The live sharpness meter in the photo panel tells you when you've found it.

Privacy & data

Everything lives in your browser's localStorage: inventory, caches, settings, API keys. Nothing is transmitted anywhere except the AI lookup requests, which go directly from your browser to the provider you selected. Clearing browser data clears the tool — export regularly.

Roadmap

• MCP server extraction of the lookup core (shared cache across workstations, server-side keys, CORS-free integrations: Icecat, eBay Browse, FCC ID)
• Cache export/import for backup and team sharing
• Two-photo (front/back) identification for dense modules

License

MIT — see LICENSE.

Built by a working ITAD technician to solve a real bench problem, with AI-assisted development. The best feature requests came from production use.