Agentic Service Layer
Govern the AI your company already uses.
The work is the combination: learn, train, moderate, and load-balance — while occupation experts keep up with the business. Alisha is only the human interface.
How it works
Learn, train, moderate, balance — and keep pace.
One loop. Experts study specific fields from your data. The control plane checks, caps, and spreads the load. Alisha talks to people.
The graph learns the field — not the leftover thread.
Work, documents, and queued study land in tenant graphs. Occupation experts keep studying their specialty from that data. Alisha speaks from what is known.
Human interface
Alisha talks to people. Experts own the field.
A.L.I.S.H.A. — AI Layer Interactive Service Human Agent. One voice per company. The specialists stay behind her.
She answers. She is not the specialist.
Alisha is the human interface — the only voice staff hear. Occupation experts study specific fields behind her. Research APIs never talk to people.
Occupation experts
They study the field. Alisha is the interface.
Expert agents keep learning very specific occupations from your data, validate what they find, and deepen without writing new code. They never speak to staff.
One expert per occupation — not a generalist.
Specialties are fields of study (knowledge graphs, tax policy, support operations), not a leftover chat topic. An expert is minted when that field shows up often enough, or when you ask them to study it.
Control plane
What you govern
Spend, safety, and rules — without handing the company to a vendor console.
Spend and activity
See which team, user, and agent spent what — while work is still in flight.
- Live operations: latency, errors, driver health
- Agent activity rollups by company and user
- Costs and errors without opening a vendor console
Developers
Built to sit in front of your stack.
Headless control plane. Point existing agent apps, MCP clients, and LLM SDKs at ASL — download the client and keep the UI you have.
SDK
Python and TypeScript clients — REST, MCP, and wrap helpers. Download and drop into the app you already have.
- Python is stdlib-only (pip install . from the zip). TypeScript is one fetch module.
- AslMiddleware.complete replaces an LLM call. wrap_api records outbound helpers, then they still run.
from asl_sdk import AslClient, AslMiddleware
client = AslClient("http://localhost:8080")
client.login("user", "user")
print(client.query("What changed?")["answer"]["text"])