History

chrischristiansen-glitch dc7fc0aa79 docs+fix(CG6): mark CG6 done in ROADMAP, fix Compute Engine duplicates in billing_agent, add GCP-only hosting principle		2026-06-10 00:26:13 +02:00
..
__init__.py	feat(ml): add ML layer — Dask DAG, Parameter Server state, telemetry, AGENTS.md [ML-1]	2026-05-26 01:13:19 +02:00
anomaly_detector.py	feat(cg6): publish anomalies to Pub/Sub topic	2026-05-27 17:29:22 +00:00
billing_agent.py	docs+fix(CG6): mark CG6 done in ROADMAP, fix Compute Engine duplicates in billing_agent, add GCP-only hosting principle	2026-06-10 00:26:13 +02:00
README.md	feat(ml): add ML layer — Dask DAG, Parameter Server state, telemetry, AGENTS.md [ML-1]	2026-05-26 01:13:19 +02:00
state_store.py	feat(ml): add ML layer — Dask DAG, Parameter Server state, telemetry, AGENTS.md [ML-1]	2026-05-26 01:13:19 +02:00
task_graph.py	feat(ml): add ML layer — Dask DAG, Parameter Server state, telemetry, AGENTS.md [ML-1]	2026-05-26 01:13:19 +02:00
telemetry.py	feat(ml): add ML layer — Dask DAG, Parameter Server state, telemetry, AGENTS.md [ML-1]	2026-05-26 01:13:19 +02:00

README.md

OSVauco ML Layer

ML-laget gir OPAX-agentene parallell kjøring, delt state-kontroll og strukturert telemetri. Implementert i tre faser.

Arkitektur

OPAX /run-endepunkt
       │
       ▼
 build_agent_dag()          ← Dask DAG (task_graph.py)
       │
  ┌────┴────┬──────────────┐
  ▼         ▼              ▼
agent_A  agent_B  ...  agent_N   ← kjøres parallelt
  │         │              │
  └────┬────┴──────────────┘
       ▼
 AgentStateStore             ← Parameter Server state (state_store.py)
       │
       ▼
 telemetry JSONL              ← rådatagrunnlag for ML-2 og ML-3

Filer

Fil	Ansvar
`task_graph.py`	Dask DAG — parallell agent-scheduling
`state_store.py`	Thread-safe PS-inspirert state store
`telemetry.py`	Strukturert JSONL-logging per agent-kall
`__init__.py`	Pakke-eksporter

Scheduler-valg

Miljø	`scheduler`-parameter
Dev / unit-test	`'synchronous'`
Lokal multi-tråd	`'threads'`
Cloud Run (lett)	`'threads'`
Full skala	`'distributed'` (krever Dask cluster)

Fase-oversikt

Fase	Innhold	Status
ML-1	Dask DAG + PS state + telemetri	✅ Implementert
ML-2	Inkrementell læring (dask-ml) + ASHA-tuning (Ray Tune)	🔜 Neste
ML-3	XGBoost feature importance + PBT	🔜 Fremtidig

Bruk

from ml import build_agent_dag, execute_dag, get_store, log_agent_call

# Bygg og kjør DAG
tasks = build_agent_dag([billing_agent, rag_agent], payload)
results = execute_dag(tasks, scheduler="threads")

# State store
store = get_store()
store.push("billing-agent", "last_cost_usd", 12.5)
cost = store.pull("billing-agent", "last_cost_usd")

# Telemetri
log_agent_call(
    agent_id="billing-agent",
    input_payload=payload,
    output=result,
    model_used="gemini-2.0-flash",
    mode="standard",
    duration_s=0.4,
    success=True,
)