Prototype 05 · The Colony

SeeStocks
does not crawl.
It pollinates.

A multi-agent AI suite that turns fragmented merchant catalogs into a unified, semantically-searchable commerce graph. At its core, a swarm of autonomous agents — internally called abejas — continuously ingests, enriches, classifies and publishes product data at scale.

QWEN3-VL-32B CLIP EMBEDDINGS PGVECTOR · HNSW GOOGLE CSS META CATALOG
Indexed SKUs
793K+
Precomputed rankings
1.5M
Semantic latency
<15ms
Active agents
06
// 02 · The Fleet

The agent fleet operates as a coordinated colony.

Each abeja owns a single domain. Together they keep the catalog graph alive, accurate, and indexed.

🐝 [01]

Vidente

vision · multimodal

Vision-language analysis on product imagery via Qwen3-VL-32B. Reads photographs the way a buyer's eye reads a shelf.

🐝 [02]

Celestina

deduplication · matchmaker

Clusters duplicate offerings across merchants using SKU/EAN signals. Turns five listings of one product into one canonical entity.

🐝 [03]

Embeddings

vectorization · CLIP

Generates dense vector representations through CLIP for semantic retrieval. Every product becomes a point in a navigable space.

🐝 [04]

Escriba

copy · validated LLM

Drafts SEO-optimized titles and descriptions through a validated LLM generation pipeline. Voice-consistent, hallucination-checked.

🐝 [05]

Catálogo

taxonomy · GPC alignment

Enriches taxonomies against Google Product Categories. Ambiguous merchant labels get pinned to a stable industry standard.

🐝 [06]

Sync

distribution · CSS · Meta

Propagates inventory and metadata to Google CSS and Meta catalogs. Whatever the colony produces, the open web sees within minutes.

// 03 · The Hive

Engineering depth, visible at every layer.

HNSW-indexed pgvector store

Custom-tuned graph index serving sub-15ms semantic search across 793K+ products. The retrieval brain of the colony.

Hybrid GPU / cloud topology

On-premise GPU clusters for inference, cloud for orchestration. Cost-efficient where it matters, elastic where it should be.

Atomic ranking refresh

1.5M precomputed orderings served from atomic refresh tables. The user sees a fresh ranking; the system never sees a half-written one.

Immutable audit trail

Every agent action is logged, attributable and replayable. No black box — only inspectable behaviour.

Docker-native deployment

Reproducible across environments. Every release is a container; every container is signed.

Self-service merchant portal

OAuth onboarding, catalog ingest, real-time enrichment status. Merchants connect once and watch their listings improve.

/// /// /// /// ///

SeeStocks isn't a marketplace.
It's the AI infrastructure layer
that makes any merchant catalog
intelligent, discoverable, and monetizable.

The team ships production-grade agentic systems daily — turning bleeding-edge LLM capabilities into measurable revenue.