Furniture Valuation CV Pipeline

The problem

A resale marketplace needed an automated system to evaluate furniture from images and listing data at scale. The business challenges were specific and commercially real:

• Is this item worth processing? — classify whether furniture is valuable or undesirable based on visual features
• What's it worth? — predict resale value from image + listing data
• Is this a scam? — detect duplicate images that indicate fraudulent listings
• Can it scale? — process hundreds of listings per minute from bulk CSV uploads via an API

This was not a notebook experiment. It was a production ML system with real operational requirements.

My role

I worked as the ML/AI engineer responsible for the computer vision pipeline, model architecture, and deployment infrastructure. The work covered training, inference, API design, and production scaling.

What I built

Quality classification

Built a visual classification system that determines whether furniture is valuable (worth resale processing) or undesirable (damaged, low-quality, or not worth the logistics). This required:

• Feature extraction from listing images
• Training on labeled quality datasets
• Balancing precision and recall for business-appropriate thresholds
• Handling varied image quality, angles, and lighting conditions

Resale value prediction

Developed a pricing model that combines visual features with listing metadata to estimate resale value. The model accounts for:

• Condition signals visible in images
• Listing data (brand, dimensions, original price where available)
• Market positioning for different furniture categories
• Confidence scoring to flag uncertain predictions

Duplicate-image scam detection

Built a duplicate/near-duplicate image detection system to catch fraudulent listings that reuse photos from other sellers. This involved:

• Image fingerprinting and perceptual hashing
• Similarity scoring across the listing database
• Threshold tuning to catch scams without false-flagging legitimate relisted items

Make/model/trim detection

Worked on fine-grained attribute prediction — identifying furniture make, model, and style characteristics from images. This is a harder classification problem than binary quality assessment because it requires:

• Visual feature discrimination between similar-looking products
• Structured attribute output rather than single-label classification
• Business-relevant labeling aligned with resale categories

Bulk processing and API deployment

The system was deployed as a containerized API capable of handling production throughput:

• FastAPI inference service — handles individual and batch prediction requests
• CSV bulk processing — ingests spreadsheets of listings, processes images, and returns enriched output
• Dockerized deployment — containerized for consistent production serving
• Scaling architecture — designed to handle 300+ listings per minute
• Backend integration — clean API contracts for downstream systems to consume predictions

Architecture

The pipeline follows a multi-stage approach:

1. Ingestion — listings arrive via API (single or CSV bulk), images are fetched and preprocessed
2. Classification — quality assessment determines if the item is worth processing further
3. Prediction — resale value estimation combines visual and metadata features
4. Fraud check — duplicate-image detection flags suspicious listings
5. Attribute extraction — make/model/style attributes are predicted for qualified items
6. Output — enriched listing data with predictions, confidence scores, and flags is returned via API

Each stage can be independently updated — new models can be swapped in without changing the API contract.

What this project proved

This is one of the strongest examples of ML engineering growth in my portfolio:

• From app development into ML-powered products — this wasn't wrapping an API; it was building the actual predictive system
• From prototype to production — Docker deployment, backend integration, bulk processing, and scaling to hundreds of listings per minute
• From generic ML into business-specific modeling — furniture valuation is business logic wrapped in ML: determining processing worthiness, estimating value, detecting fraud
• From one prediction task into multi-task systems — quality classification, price prediction, duplicate detection, and attribute extraction in one pipeline
• From model work into system integration — training, inference serving, API design, and operational handoff

Outcome

Delivered a production-ready computer vision pipeline that automated what would otherwise require human evaluators to manually assess hundreds of listings. The system handles classification, pricing, fraud detection, and attribute extraction at scale — demonstrating that my ML work goes beyond prototypes into commercially deployed predictive systems.