Systems and architecture

1Where people and models work

HelixHumans and model seats in one room: chat, voice, Teams.

2Where work is governed

AgentOSGoverns the engineering: work orders, review, cost.Agent SwarmRuns the agents in sandboxes, driven by events.

3What agents reach through

FAFOHive475 tools behind one endpoint, memory and code intelligence inside.FAFO Core and Function ForgeOne function becomes an orchestrated, governed tool.

4What the work runs on

PostgreSQL FabricOne write path, sized and admitted from the database's own limits.Inference FabricToken-aware brokering, saturated runtimes, a CUDA analytics engine.ActiveKVConversation history in host memory, the GPU as a bounded working set.

Across every plane

Capabilities modelA tool catalog compiled into a small model with a deterministic decoder.Auth and access planeCommercial assistants admitted to my tools on my terms.

Numbered in the order work travels

AgentOS
A governed work-order control plane for teams of AI engineering agents
8 stories39 technologies
Agent Swarm
A self-hosted, multi-host runtime where tools are the platform and agents are compositions
4 stories40 technologies
Helix
Humans and several AI minds in one room, and one colleague who can be reached wherever the work is: in team chat, by voice, and in a Teams call
17 stories52 technologies
FAFOHive
The MCP hub every agent in my estate connects to: memory and code intelligence, hosted tools, and a governed tool gateway
8 stories60 technologies
FAFO Core and Function Forge
The execution substrate under everything else: one function definition becomes an orchestrated, budgeted, agent-callable tool
12 stories56 technologies
PostgreSQL Fabric
One governed write path for a GPU-scale estate: batches sized from the database's own limits, workers that follow the queue, admission that reads the database's pressure, and the heavy work run inside the database next to its indexes
9 stories11 technologies
Inference Fabric
A throughput-first GPU inference pipeline on consumer NVIDIA hardware: broker, saturated runtime, completion, vector store
39 stories53 technologies
ActiveKV
An inference-memory layer for vLLM: the conversation's history in host RAM, the GPU as a bounded working set
10 stories26 technologies
Capabilities model
A 0.6-billion-parameter model that memorises an agent's tool catalog, with the data, decoder, serving and evaluation around it
13 stories42 technologies
Auth and access plane
How commercial AI assistants and my own agents reach a private tool estate: OAuth, token validation, edge exposure and an operations layer
1 story36 technologies