The workbench under RaboAurora
Build the data, models and agents the platform runs on.
Forge is where builders work: shape a data product and its contract, train a model on governed data, compose an agent on a canvas, orchestrate a fleet of them, prove it with a harness, and schedule the whole thing.
D
Model the data
Data products
Wire a pipeline, declare the schema and contract, attach quality rules and publish serving ports.
4 products · 31 checks
M
Adapt
Model builder
QLoRA training on Unsloth — 2× faster, 68% less VRAM. Governed datasets in, adapters and GGUF out.
4 runs · onb-lora-v4
B
Compose
Agent builder
Drag-and-drop canvas for wiring models, retrievers, tools, guardrails and human gates into an agent.
7 blocks · v4 draft
O
Coordinate
Agent orchestrator
A main agent that spawns sub-agents, merges their answers and escalates when confidence drops.
4 agents · 3.4k sessions
H
Prove
Agent harnessing
The evaluation gate: functional, grounding, safety, red team and compliance suites with regression tracking.
714 scenarios · 96%
C
Ground
Agent context
Sources, chunking, retrieval and the token budget — what the agent actually knows when it answers.
3.5k chunks · recall 94%
S
Automate
Scheduler
Cron, event and webhook triggers for reindexing, evaluation runs, backlog sweeps and fine-tune refreshes.
6 schedules · 312 runs today
A
How it connects
Forge publishes reusable work, and the harness feeds Alibi
Data productspublish straight into the catalog
Adapterstrained models become Forge runtimes
Agentscanvas graphs publish into Forge
Evidenceharness results land in Alibi
Scheduleskeep context and evals fresh