Inference Architecture
The ind.ai Inference Plan
Router-led, cost-aware, outcome-driven
ind.ai does not default to the most powerful model. It routes every request through a scoring system that selects the cheapest viable execution path — escalating to proprietary models only when task complexity, risk, or quality requirements demand it.
Inference Thesis
Three principles
Open-source is the mass-market layer
OSS models handle the majority of inference volume. They are fast, cheap, and improving rapidly. Most user requests never need proprietary models.
Proprietary is the escalation layer
Flagship models from frontier labs are used selectively — when reasoning depth, risk, or quality thresholds exceed what OSS can deliver.
The router decides the path
A scoring system evaluates every request across multiple dimensions and selects the cheapest viable execution class. No model is default.
Escalation chain — always start left, move right only when required
Execution Classes
Five execution classes
Every inference request is classified into one of five execution classes. The router selects the lowest viable class.
Deterministic
No model call. Direct lookup, API response, or cached result.
OSS-Small
Lightweight open-source model. Fast, cheap, high-volume.
OSS-Mid
Larger open-source model. Stronger reasoning, still cost-efficient.
Proprietary-Mini
Smaller proprietary model. Better quality where OSS falls short.
Proprietary-Premium
Flagship proprietary model. Reserved for high-stakes, high-complexity tasks.
Task Mapping
Which tasks go where
Representative mapping — actual routing is dynamic and score-based.
Tier Access
Which users get which inference
Tier determines model class ceiling, not just usage volume.
Free
- Class 0–1 unlimited
- Class 2 with daily limits
- No Class 3–4 access
- OSS models only
- Standard response quality
Pro
- Class 0–2 unlimited
- Class 3 with generous limits
- Class 4 with capped usage
- Proprietary model access
- Priority queue
Enterprise
- All classes unlimited
- Custom model routing rules
- Dedicated inference capacity
- Private model deployment
- SLA-backed response times
Router Logic
Model router scoring dimensions
The router evaluates each request across these dimensions to select the optimal execution class.
Task type
Classification of the incoming request against known task patterns.
Ambiguity
How unclear or multi-interpretable the request is.
Risk
Potential impact of incorrect output — financial, legal, health.
Reasoning depth
Number of logical steps required for accurate completion.
Language quality
Required fluency, tone precision, and multilingual demands.
User tier
Free, Pro, or Enterprise — determines available model classes.
Workflow sensitivity
Whether the task involves irreversible actions or approvals.
Cost budget
Remaining cost budget for the user's current billing period.
India Context
India-specific considerations
Multilingual routing
Language detection influences model selection. Bhashini-compatible OSS models handle most Indic language tasks; proprietary models are escalated only for complex multilingual reasoning.
Frugal inference
India's cost sensitivity demands aggressive optimisation. The router defaults to the cheapest viable path and avoids proprietary models for routine tasks.
Escalation logic
High-risk tasks (financial, health, legal) trigger mandatory escalation regardless of user tier. Safety overrides cost optimisation.
Cost Discipline
Cost-control rules
Operational discipline that keeps inference economically sustainable at scale.
Connected Architecture
How this connects to the Starter Agent Fabric
The inference plan defines which models run. The Starter Agent Fabric defines how agents are structured. Together, they form ind.ai's execution architecture: a shared orchestration layer that routes requests through the cheapest viable inference path, wrapped in identity, memory, and tier-based access control.
Free users get fast, cost-efficient agents powered by deterministic and open-source inference. Pro users unlock proprietary model access for complex tasks. Enterprise users get custom routing, dedicated capacity, and SLA-backed execution.
ind.ai should not be model-branded. It should be router-led, cost-aware, and outcome-driven.
The inference plan ensures that every request gets the right model — not the most expensive one. This is what makes India-scale AI agents economically viable.
Related in India AI