imm-spec-agent

name

imm-spec-agent

description

Use EVERY TIME someone wants a spec sheet, research document, or freight plan for a used plastic injection molding machine. Trigger on brand+model (KraussMaffei CX 160, Engel Victory, Haitian Mars, Arburg Allrounder), an equipt.com/exapro.com/machinepoint URL, a photo of a machine, or phrases like "build me a spec sheet", "research this IMM", "how do I ship this machine", "step deck or RGN", "how many trucks", "create a one-pager for this used injection molder". Produces a premium dark-theme PDF (matching La Crown's GEO report aesthetic) white-labeled with the user's brand colors pulled from their website on first run. Includes hero photo + multi-photo gallery, hyperlinked listing button on cover, asking price, full technical specs in metric AND imperial, freight logistics with trailer/permit/escort/rate recommendations, blueprint links, and source citations. OEM-first, reseller fallback. All embedded photos auto-strip metadata silently.

IMM Spec Agent (Premium White-Label)

End-to-end research agent for used plastic injection molding machines. Output is a branded PDF spec sheet matching the La Crown GEO report aesthetic — dark theme, Playfair Display serif headlines, two-tone accent words, gold/orange critical numbers. Brand identity is pulled from the user's website on first run.

When this fires

• Brand + model: "KraussMaffei CX 160", "Engel victory 200/120", "Haitian Mars II 250"
• A listing URL: equipt.com, exapro.com, machinepoint.com, surplusrecord.com
• A photo of a machine (read nameplate or visual ID)
• Phrases: "build me a spec sheet", "research this IMM", "give me the freight plan", "how do I ship this"

First-run flow (every invocation)

1. Check for cached brand at /mnt/user-data/uploads/.imm_brand.json. If exists, load and skip to step 3.

2. No cached brand → ask for the user's website ONCE:

   "Before I build the spec sheet — what's your company website? I'll pull your brand colors and contact info so the PDF is white-labeled. (Or paste 'skip' for generic branding.)"

   When given a URL:

   • Run python3 scripts/brand_extractor.py <URL> /mnt/user-data/uploads/.imm_brand.json
   • Confirm the extracted colors with the user before saving.
   • Save final brand to cache.

   On "skip" → use generic default (gold/orange on dark theme).

3. Proceed to spec sheet workflow below.

What to ask before building (or during)

The agent uses what's provided and prompts for the rest at the end — never blocks the spec sheet to demand inputs.

Input	When to ask	Goes where
Hero photo	At end if missing	Cover page
Additional photos (nameplate, control panel, injection unit, etc.)	At end if missing — encourage nameplate especially	"Machine Gallery" page
Listing URL on reseller's site	At end if missing	Gold "View Live Listing →" button on cover
Asking price	At end if missing	Italic Playfair price on cover (next to listing button)
Production hours	At end if missing — this is the #1 used-IMM buyer question	Cover stat chip + Identification table
Condition rating	Fallback if hours unknown (Excellent / Good / Fair / As-Is)	Cover stat chip (if no hours) + Identification table
Inventory ID / stock number	Optional	Footer ribbon if provided

Closing prompt (if any are missing)

After delivering the first draft, ALWAYS end with something like:

"To upgrade this from research draft to customer-ready, drop me: • A nameplate close-up (upgrades all the LOW-confidence fields to HIGH) • 2-3 more shots — control panel, injection unit, power pack • Production hours (or condition rating if hours aren't logged) • The listing URL on your site if you have one • Asking price (or 'price on request') Send any combo and I'll regenerate."

Photo handling — ALWAYS auto-scrub

Every photo embedded in the PDF gets metadata-stripped automatically before embedding. Use prep_photos.py:

from prep_photos import prep_photos

# User dropped raw photos at /mnt/user-data/uploads/
raw = ["/mnt/user-data/uploads/hero.jpg",
       "/mnt/user-data/uploads/nameplate.jpg",
       "/mnt/user-data/uploads/control.jpg"]
clean = prep_photos(raw, "/tmp/spec_build/clean")

spec["photo_paths"] = [str(p) for p in clean]
spec["photo_captions"] = {
    str(clean[1]): "Nameplate Close-up",
    str(clean[2]): "Operator Side / Control",
}

The first photo in photo_paths is the hero (cover page). Everything after becomes the gallery. Captions are optional dict mapping path → label.

No telling the user it's happening. The privacy line ("metadata-stripped") on the gallery page is the only visible signal.

Spec workflow

Step 1 — Identify the machine

• URL given → web_fetch it
• Brand + model → start here
• Photo → read nameplate / visual ID. Customer-applied asset tags (e.g. "A60") are NOT OEM models.

Step 2 — Pull OEM data first

web_search: "<Brand> <model> specifications dimensions weight"
web_search: "<Brand> <series> brochure filetype:pdf site:<OEM domain>"

See references/data-sources.md for OEM domains.

Step 3 — Cross-reference with reseller listings

Equipt, Exapro, Maszyneria, MachinePoint, Surplus Record.

Step 4 — Run conversions + freight engine

from convert import format_full_spec
from logistics import recommend_freight

formatted = format_full_spec(raw)
plan = recommend_freight(machine_l_mm=..., machine_w_mm=...,
                         machine_h_mm=..., net_weight_kg=...)

Step 5 — Build the branded PDF

import json
from build_pdf import build_spec_pdf

with open('/mnt/user-data/uploads/.imm_brand.json') as f:
    brand = json.load(f)

spec = {
    "type": "Plastic Injection Molding Machine",
    "manufacturer": "Krauss-Maffei",
    "model": "C-Series · 300 Ton",
    "year": "2003",
    "serial": "12345",
    "production_hours": 42500,              # int or str — appears as cover chip + identity row
    "condition": "Good — Working Condition", # fallback chip if hours missing
    "description": "...HTML allowed for <b>bold</b> + <br/>...",
    "specs_formatted": format_full_spec(raw),
    "logistics": plan.to_dict(),
    "photo_paths": clean_photo_paths,       # list, all auto-scrubbed
    "photo_captions": {path: caption, ...}, # optional
    "listing_url": "https://reseller.com/listing/12345",  # → cover button
    "asking_price": "$48,500",              # → cover price callout
    "blueprint_links": [...],
    "sources": [...],
    "confidence_notes": "...",
}

build_spec_pdf(spec, "/mnt/user-data/outputs/<brand>_<model>_spec.pdf",
               brand=brand)

Critical rules

• First run always asks for the website. Don't skip. Don't guess colors.
• Brand is cached in /mnt/user-data/uploads/.imm_brand.json — subsequent runs use it without re-asking.
• OEM data first, reseller second. When they conflict, prefer OEM.
• Both unit systems on every numeric field. Never just metric, never just imperial.
• Mark confidence honestly. HIGH / MEDIUM / LOW per field.
• Run web_search every time. Used-IMM specs you "remember" may be wrong revision.
• Don't invent serials, years, or hours. "n/a" or "TBC from nameplate".
• Photos are optional. Build whatever you have. Prompt for more at the end.
• All embedded photos are auto-scrubbed. No exceptions.
• The PDF is always the deliverable, even for quick questions.

Resetting / overriding the brand

"reset brand" or "change my brand colors" → delete .imm_brand.json and re-run first-run flow.

"build this for [Other Company]" → temporarily override company_name for one-time use without touching the cache.

File map

• scripts/brand_extractor.py — pulls company name, logo, colors, contact from a website
• scripts/prep_photos.py — strips metadata from photos (used silently before embedding)
• scripts/convert.py — unit conversions (metric ↔ imperial)
• scripts/logistics.py — trailer/permit/escort/rate engine
• scripts/build_pdf.py — HTML + WeasyPrint PDF generator (premium dark theme)
• references/data-sources.md — OEM domains + reseller aggregators by brand
• references/logistics-rules.md — freight rules, trailer types, brand-specific shipping

Anti-patterns

• ❌ Asking for photos upfront and refusing to build without them → build with whatever's provided, prompt at end
• ❌ Including raw photos with GPS/timestamps in the PDF → always run prep_photos first
• ❌ Hardcoding any reseller's branding into the default → use generic gold/dark when no brand is set
• ❌ Skipping the listing URL when one was provided → always render the gold button on the cover
• ❌ Skipping the brand confirmation step → always show extracted colors before caching
• ❌ Reusing a Premier-branded sample for other resellers → samples are always GENERIC unless a specific brand is requested