Multi-agent orchestration platform
The core service: runs ordered teams of LLM agents, pauses for a human answer and resumes where it left off, imports and exports teams as portable bundles, and serves each team its own model catalogue.
python · fastapi · async postgres · litellm
01
Bring-your-own-key LLM billing
Every run is paid by the owning team's encrypted credential, resolved once at start and again on resume, so a rotated key never changes a run mid-flight. It fails loudly instead of falling back, and no secret reaches a log, error, span or response.
fernet · settings validation · audit log
02
Access control that doesn't leak
Membership, ownership and super-admin rules for people, services and agents. A missing resource and a forbidden one are refused the same way, so the API never reveals which ids exist.
entra id · jwt · m2m roles · rbac
03
Agent tool service
Tools for agents behind HTTP: self-describing manifests, per-tool timeouts, API-key clients, and toolkits that fail alone without taking the service down.
fastapi · sqlalchemy · web search · mcp
04
Semantic catalog vectors
A batch pipeline that embeds thousands of catalog values into pgvector, so agents match by meaning instead of exact text. Operators launch it from a template; secrets stay in the vault.
apache beam · dataflow · pgvector · secret manager
05
Product categorization
Classifying large product catalogs with a small open model, run as reproducible batch jobs.
gemma · batch inference
06
An engineering workflow with Claude Code
Custom subagents that keep the work honest: a strict test-first implementer, an architecture reviewer, a code-bloat reviewer, and one that logs each day's commits to Jira. Spec-driven skills take a feature from intake to tasks.
claude code · subagents · tdd
07