Oraclous is built in three composable layers. Layer 1 stores your domain knowledge. Layer 2 gives agents the tools to act on it. Layer 3 runs the autonomous fine-tuning loop on top.
The autonomous fine-tuning operation
The infrastructure agents run on
The data foundation everything builds on
18 specialist agents that automate the complete fine-tuning lifecycle — from knowledge graph analysis to model deployment and drift monitoring.
Every fine-tuning cycle follows a deterministic 10-stage pipeline: Connect → Structure → Analyze → Research → Curate → Select → Train → Evaluate → Deploy → Monitor. Each stage is owned by a dedicated agent. The loop re-enters at Stage 3 when the Monitor Agent detects drift above threshold.
HITL approval UIs appear at every critical decision point — dataset review, strategy selection, training launch, deployment promotion. Operators can inspect, modify, or override agent decisions before proceeding. No stage advances without explicit approval unless you configure fully autonomous mode.
The Strategy Agent (Stage 6) chooses the optimal training technique based on dataset characteristics, target capability, and compute constraints. Supported methods: SFT (supervised fine-tuning), LoRA, QLoRA, full fine-tune, DPO (direct preference optimization), and ORPO. Hyperparameters are generated, not guessed.