ONDA: 6 Levels of AI Self-Improvement
What is ONDA?
ONDA stands for Observation, Navigation, Decision, Action. It is the self-learning framework at the core of Plugy. Rather than treating AI optimization as a single knob to turn, ONDA breaks it into six distinct levels, each building on the one below. The system climbs these levels automatically, improving itself with every conversation.
Level 0: Raw Input Processing
The foundation. At this level, the system learns to correctly parse and understand incoming messages. This includes language detection, intent classification, entity extraction, and context assembly. A message like “my server 192.168.1.5 is down since morning” needs to be decomposed into intent (report issue), entity (server IP), and temporal context (since morning).
Level 0 optimization means fewer misunderstood messages and more accurate routing.
Level 1: Persona Optimization
How the AI sounds matters. Level 1 optimizes tone, formality, response length, and communication style. A hosting company’s support should sound different from a fashion brand’s chatbot. The system learns from operator corrections and customer reactions which persona works best for each context.
Level 2: Business Rules Refinement
Every company has policies: refund windows, escalation paths, SLA thresholds. Level 2 continuously refines how the AI applies these rules. When an operator overrides the bot’s decision, the system learns which rule was wrong and adjusts. Over time, the AI’s policy application converges with how your best operators handle edge cases.
Level 3: Knowledge Retrieval Optimization
Having a knowledge base is not enough — the AI needs to find the right article at the right time. Level 3 optimizes retrieval: which documents are most relevant, how to chunk information, when to combine multiple sources, and when to admit the answer is not in the knowledge base rather than hallucinate.
Level 4: Model Routing
Not every question needs the same AI model. Simple FAQs can use fast, lightweight models. Complex technical questions benefit from more capable (and expensive) models. Level 4 learns which model performs best for which scenario, optimizing both quality and cost automatically.
Level 5: Self-Observation (The Strange Loop)
This is where ONDA becomes unique. Level 5 is the meta-level: the system observes its own optimization process. It monitors whether changes at Levels 0-4 actually improved outcomes. If a Level 1 persona change decreased customer satisfaction, Level 5 detects this and triggers a rollback.
The mathematical guarantee is critical: every optimization step is validated against a quality metric before being committed. This creates a monotonically non-decreasing quality curve. The system can only get better or stay the same — never degrade.
Why Six Levels Matter
Most AI support tools optimize at one or two levels — typically knowledge retrieval and maybe tone. ONDA’s six-level approach means improvements compound. Better input processing feeds better persona selection, which feeds better rule application, which feeds better retrieval, which feeds smarter model routing. And Level 5 ensures the whole stack moves in the right direction.
This is not incremental improvement. It is systematic, guaranteed progress — built into the platform from the ground up.