Collection layer
Chrome extension (Manifest v3) captures product pages from major Korean commerce sources and posts to the collect API; per-platform identifiers drive group-level deduplication; 2.19M collection jobs, 7.88M items to date.
CODOT · WONJU, KOREA · COMMERCE AUTOMATION R&D
Commerce automation, engineered as one pipeline.
Codot builds the AI that turns product sourcing, content generation, marketplace listing, and order operations into a single data pipeline. Our platform, SellingMate, is a web + Chrome-extension SaaS with 465 API endpoints, 7 job queues, and a stack of AI engines — salient-object segmentation, neural rendering, retrieval-grounded structuring, and demand forecasting — running in production at more than a million products a month.
In production · 1M+ products processed every month
The problem
A marketplace listing is 20–30 structured fields, each governed by rules that differ per marketplace and change without notice. Getting one wrong hides the listing or suspends the account. At hundreds of products a day this is not a labor problem — it is a data-structuring and rule-learning problem, and that is the problem Codot is built to solve.
Structured data in, compliant listings out.
- Product records collected
- 7.9M
- Products structured by AI
- 4.07M
- Marketplace listings mapped
- 2.19M
Architecture
Four layers, one data flow.
Structuring layer
Transformer-based attribute extraction grounded by retrieval over the marketplace rule corpus (RAG on ChromaDB) turns titles, options, and specs into a listing-ready schema; CJK attribute names are mapped through a learned display dictionary; a GTIN cache validates check digits and string similarity. Average 2.16 s per item.
Listing layer
Marketplace rules are stored as data (7,130 blocked keywords, 168 blocked categories) and embedded for retrieval; a hierarchical category classifier proposes categories and self-corrects from rejection feedback; preflight validates every field, replaces banned terms, and enriches from 8 external sources. 2.19M listing mappings, 1.78M successful.
Operations layer
Adaptive scheduler syncs orders every 60 minutes and returns on a configurable cycle; 30,583 return requests handled; 2.24M double-entry ledger rows; team roles and permissions.
Celery on Redis · 7 queues (default, uploads, removebg, vendor_recheck, meta_preflight, collect, deletes) · 14 routed task types · late-ack, prefetch 1 · adaptive scheduler with stale-job sweeps
AI engines
Every engine ships to real accounts the day it works.
| Engine | Modality | Stage | Next milestone | Volume to date |
|---|---|---|---|---|
| Background removal | Salient-object segmentation, U²-Net family cascade | Production (Aug 2025) |
On-prem GPU inference (2027) | 672K images, peak 110K/month |
| Attribute extraction & structuring | LLM + retrieval-grounded rules (RAG / ChromaDB) | Production (Jul 2025) |
Multilingual attribute schema | 4.07M products |
| Listing preflight & category mapping | Hierarchical classifier + rule engine | Production |
Self-correcting category mapping v2 | 191K checks, 66K enrichments |
| Generative image studio | Diffusion models + NeRF view synthesis | Production (2025) |
Batch generation, brand presets | 4,337 images, 94 templates |
| Sales prediction & recommendation | LSTM sequence models + gradient-boosted ranking | Production (2025) |
Recommendation API for partners | Trained on 4.07M products / 213K orders |
| 3D product twins | Neural rendering (3D Gaussian splatting) | In development |
Multi-view capture → 360° product assets (2027) | Multi-view capture pipeline in progress |
| Commerce data platform | Data platform | In development |
Partner data API (2027) | Schema and access model in design |
Computer vision
Salient-object segmentation cascade
A U²-Net family cascade (u2netp → silueta → u2net) with in-process session caching, followed by OpenCV post-processing: badge suppression, alpha-matte edge refinement, white-background compositing, and marketplace-spec 1,000 px resize. 672K images, peak 110K a month.
Rules engine
Retrieval-grounded listing preflight
Marketplace rules are embedded and retrieved (RAG over ChromaDB) so the model structures each field against the current rule set; a hierarchical category classifier proposes categories, validates them against 168 blocked nodes, and self-corrects from rejection feedback. 191K preflight runs, 66K enrichment jobs.
Generative AI
Neural rendering & diffusion studio
Product photos become new angles through neural radiance field (NeRF) view synthesis and diffusion-based background generation; template-conditioned prompts fix subject, lighting, and 3/4 view, and 94 detail-page templates assemble the output. 81 batch jobs, 4,337 images.
Prediction
Demand forecasting & ranking
LSTM sequence models and gradient-boosted ensembles over 4.07M products and 213K orders, with engineered price, category, and listing-age signals, learning what listed products actually sold. In production since 2025, recommending what to list next; accuracy improves as sales data accumulates.
Data
A closed loop from listing to learning.
Every listing, order, and return flows back into the same database, so the forecasting models train on ground truth the platform itself produced — a closed data loop that also feeds the rule corpus and the category classifier.
- products
- 4.07M
- listing mappings
- 2.19M
- orders
- 213K
- ledger rows
- 2.24M
- blocked keywords / categories
- 7,130 / 168
- detail templates
- 94
| Month | Collected | Listed | Background-removed |
|---|---|---|---|
| Mar 2026 | 1.03M | 550K | 94K |
| Apr 2026 | 1.17M | 520K | 93K |
| Jul 2026 | 1.25M | 930K | 110K |
| Aug 2026 | 0.99M | 800K | 67K |
Research & recognition
Research first, then product.
Best Paper Award, Korean Society of AI Convergence Technology
Deep-learning and time-series research by Codot's founding team.
Kaggle, top 1%
Software-defect classification with ensemble ML, by Codot's founding team.
Corporate-affiliated R&D center
Codot's research unit for commerce AI.
4.07M-product, 213K-order commerce corpus
Operational data behind attribute extraction and sales prediction; the basis for joint-research proposals.
2025 – present
- PyTorch
- Keras
- Diffusion models
- NeRF
- U²-Net
- RAG / ChromaDB
- OpenCV
- LSTM
- Flask
- Spring
- React
- React Native
- PostgreSQL
- Redis / Celery
- AWS S3
Live-account validation
Every engine is proven on real seller accounts before it is called done.
Fleet-wide upgrades
Old clients are retired automatically so the entire fleet runs the current collector within hours.
Closed data loop
Listings, orders, and returns feed the forecasting models continuously.
Infrastructure & security
Built to hold other people's catalogs.
Storage
Images on S3 with server-side encryption (AES-256 or KMS).
Database
PostgreSQL; logical backups encrypted with AES-256-CBC + PBKDF2, production guarded.
Workers
Redis-backed Celery, JSON serialization, late acknowledgement, prefetch 1, 6-hour visibility timeout.
Request security
Origin/Referer CSRF checks on state changes; restricted web/extension origins; production-only secure, HttpOnly, SameSite session cookies.
Outbound fetch
Allow-listed hosts, private/internal IP blocking, redirect and DNS pinning checks.
GPU inference
On-prem GPU cluster for the vision and rendering engines, in build-out for 2027.
Roadmap
2025
- first version on live accounts (Jul)
- background-removal cascade in production (Aug)
- generative image studio and sales prediction engine in production
- Chrome extension public release (Dec)
2026
- Codot Inc. incorporated (Jul)
- corporate R&D center (Aug)
- fleet-wide upgrade policy and recommendation API
2027
- self-correcting category model v2
- 3D Gaussian splatting product twins
- commerce data platform with partner API
- on-prem GPU inference cluster
Founding team
Sihoon Yeom — CEO, Head of Engineering
B.Eng. Digital Healthcare, Yonsei University. Three years of AI and data research (conference best-paper award, Kaggle top 1%); has run his own marketplace store since April 2025. Leads Codot's architecture and AI.
Hyebin Lee — Researcher, Infrastructure & Security
B.Eng. Advanced Materials, Hallym University. Runs cloud infrastructure and security. Former undergraduate researcher at Hallym BNDL Lab (measurement systems, quantitative experimental data).
Work with us
Building with research and integration partners.
Joint research — attribute extraction, category mapping, demand prediction on our commerce data
Integration — connect a product database or ERP through the API (465 endpoints)
Pilot — run the pipeline on a partner catalog end to end.
FAQ
Which models power the AI engines?
Segmentation: a U²-Net family cascade (u2netp, silueta, u2net) with Codot's OpenCV post-processing. Structuring and preflight: an LLM grounded by retrieval over the marketplace rule corpus (RAG on ChromaDB) plus Codot's hierarchical category classifier. Image studio: diffusion models and NeRF view synthesis under Codot's prompt-and-template pipeline. Forecasting: Codot's LSTM and gradient-boosted ranking models.
How is customer data isolated?
Every customer runs as its own tenant: team-scoped accounts, roles, and API credentials, with no cross-tenant queries anywhere in the pipeline. Product images live in S3 under server-side encryption (AES-256 / KMS); database backups are encrypted with AES-256-CBC + PBKDF2 behind production access guards. The API accepts requests only from approved web and extension origins, enforces CSRF checks on every state change, and blocks outbound requests to private networks with DNS pinning. Sessions use secure, HttpOnly, SameSite cookies.
Can we integrate through the API?
yes, 465 endpoints covering collection, structuring, listing, orders, and ledger; integration partners get a dedicated engineer.
How is the system deployed?
web console plus Chrome extension, nothing to install on the client; server side runs Flask APIs, Celery workers on Redis, PostgreSQL, and S3.