When someone takes a photo of a dusty Canon camera at a garage sale and our AI comes back with "Canon AE-1 Program, 35mm SLR, manufactured 1981-1986, estimated value $120-180 based on current condition" — what actually happened? Here's the technical story, written for people who are curious but not necessarily engineers.
The Problem
Product identification sounds simple until you try to build it. A single product can look completely different depending on angle, lighting, age, and wear. The same Canon AE-1 might be photographed in a dimly-lit thrift store, on a white background, or in a pile of other cameras. And "Canon AE-1" could mean five different variants across a 10-year production run, each with different values.
On top of identification, you need pricing — not list prices or retail prices, but actual market prices based on what comparable items sell for right now, in this condition, on this marketplace.
The Architecture
ListForge's research pipeline is built on an agent architecture. Instead of a single model that tries to do everything, we use specialized agents that each handle one part of the problem and pass results to the next.
Phase 1: Multi-Strategy Identification
Identification doesn't rely on a single method. We run multiple strategies in parallel and combine their results:
Visual Recognition — An AI model analyzes the photos to identify product type, brand visual cues, shape, and distinguishing features. This is the primary identification method and works for most items.
Text Extraction — Separately, we look for readable text in the images: brand names, model numbers, serial numbers, specifications printed on the item. A readable model number is often the single most definitive identification signal.
Barcode Lookup — If a UPC, EAN, or QR code is visible, we decode it and look it up in product databases. This gives us an exact product match with no ambiguity.
Product Knowledge Base Search — We search ListForge's own database of previously identified products using visual similarity. If we've seen this exact product before (identified and confirmed by a user), we can match it with high confidence.
These strategies produce candidate identifications that are merged and ranked. When multiple strategies agree — say, visual recognition suggests "Canon AE-1" and text extraction reads "AE-1 PROGRAM" on the top plate — confidence goes up. When they disagree, confidence drops and the system flags uncertainty.
Phase 2: Market Research
Once we know what the product is, we need to find out what it's worth. This means searching for comparable sales.
We query multiple marketplace data sources for recently sold and currently listed items matching the identification. But raw search results aren't enough — we need to filter for true comparables:
- Is this the same model and variant?
- Is the condition comparable?
- Is the sale recent enough to reflect current market conditions?
- Was the sale an outlier (damaged lot, auction fluke, charity sale)?
The AI evaluates each potential comparable and assigns a relevance score. Only high-relevance comps make it into the pricing analysis.
Phase 3: Pricing Analysis
With a set of vetted comparables, the pricing engine does its work:
- Cluster sold prices — group by condition tier (new, like new, good, fair)
- Weight by recency — more recent sales count more
- Calculate statistics — median, mean, percentiles for each condition tier
- Apply condition assessment — map the current item's condition to the appropriate tier
- Generate strategies — calculate Aggressive (25th percentile), Balanced (median), and Premium (75th percentile) price points
- Estimate sell-through — predict how long each strategy will take based on the market's velocity
Each strategy comes with a written explanation: "Priced at $95 (Balanced) based on 12 sold comparables in Good condition over the past 60 days. Median sold price was $97. Estimated 7–14 days to sell."
Phase 4: Listing Generation
The final phase takes everything learned and generates marketplace-ready listings:
- Titles crafted to include the keywords that perform best on each marketplace's search algorithm
- Descriptions that accurately represent the item, highlighting features and disclosing condition
- Category mapping to each marketplace's taxonomy
- Attribute filling for marketplace-specific required fields (item specifics on eBay, browse nodes on Amazon)
The Knowledge Base
Every item processed by ListForge contributes to a shared Product Knowledge Base (PKB). When a user confirms that "yes, this is a Canon AE-1 Program" — that identification, along with the associated images and attributes, strengthens the system's ability to recognize that product in the future.
The PKB has two tiers:
Observations — Individual data points: "this image was identified as Product X," "this item sold for $Y on marketplace Z." These are raw signals.
Canonical Products — Consolidated product identities built from many observations. A canonical product for the Canon AE-1 Program might draw from hundreds of observations across sellers, conditions, and marketplaces.
This means ListForge gets smarter over time. The more items the platform processes, the better it gets at identifying products and finding relevant comparables.
Why Evidence Matters
Every pricing recommendation links back to the specific comparable sales that informed it. This isn't just a nice-to-have — it's fundamental to the product philosophy.
Resellers are experts. They know their markets, their niches, their customers. An AI that says "price it at $45" with no explanation is asking for blind trust. An AI that says "here are 8 comparable sales, here's why I weighted these three more heavily, and here's my recommendation" is giving you information to make a better decision.
That's the difference between an AI assistant and an AI overlord. We're building the assistant.
What We're Still Improving
We're honest about where the system has gaps:
- Very rare items with no comparable sales are difficult to price accurately
- Condition assessment from photos is good but not perfect — especially for functional defects you can't see
- Category-specific nuance varies — the AI is stronger in categories where more data exists
- Speed — we're always working on making research faster without sacrificing accuracy
This is a living system that improves with every item processed, every user correction, and every model update. The AI you use today is the worst it will ever be.