Section 01 · six sensor families
Documents are not enough.
six families · each with its own ingestion pattern
Six families. Each has its own ingestion pattern and its own failure mode.
Section 02 · document parsing 2026
An AI reads the page like a person would.
documents arrive already understood, not just scanned
Old stack: scan then guess the layout. New stack: one model reads the whole page at once.
Section 03 · CDC into vector store
Change data capture - embeddings stay fresh.
Debezium 3.5 SparseVector cross-connector (Mar 2026)
For Postgres-native shops, this is logical replication - free with the database.
Section 04 · event streaming for agents
Redpanda Agentic Data Plane (Feb 2026).
streaming substrate purpose-built for agent events
Redpanda's pivot in 2026: the streaming substrate is the agent data plane.
Section 05 · schema registry
Apicurio - stores agent artifacts.
first registry to host A2A Agent Cards + MCP prompts alongside Avro/Protobuf
Confluent and Buf will follow within 12 months. Apicurio shipped first.
Section 06 · tools 2026
The OSS-first stack.
VLM-native ↑↓ legacy OCR · OSS ←→ SaaS
For self-hosted: Granite-Docling. For paid quality: Mistral OCR 3. For verticals: Reducto.
Section 07 · vollko OSS
The primitives.
· · ·
Build the AI-native firm