HOME/TECHNIQUE/Serving & Inference/Model routing & cascades

TECHNIQUE

Model routing & cascades

Serving & Inference

2APPLICATIONS
2OBSERVED OPERATORS
01

State of Practice

CROSS-VALIDATED — 4 OPERATORS

Across the deployed pool, model routing and cascades are implemented as application/platform orchestration: operators combine pre-filters, stage-specific models, fallback paths, gateways, and human review rather than relying on one model call.

Observed Practices

Use thresholded cascades so cheaper or simpler checks handle clear cases and LLMs handle only the uncertain or suspicious band.

2 of 4 operators
Kalvium LabsThumbtack

Route low-confidence or suspicious outputs to human review instead of letting the model decision stand automatically.

2 of 4 operators
Kalvium LabsThumbtack

Assign different models or model classes to different stages or roles in the workflow.

2 of 4 operators
Kalvium LabsSlack

Put multiple providers or model endpoints behind a central platform abstraction or gateway.

2 of 4 operators
GrabSlack

Use automatic fallback, spillover, or real-time rerouting when a model or endpoint is degraded, limited, or over reserved capacity.

1 of 4 operators
Slack

Log or expose routing/cascade execution details for audit, debugging, cost tracking, or future correction datasets.

3 of 4 operators
Kalvium LabsGrabSlack

Use structured outputs or schemas between cascade stages so downstream stages receive constrained, machine-readable results.

2 of 4 operators
Kalvium LabsSlack

Where Operators Converge

All four observed operators implement routing or cascading in application/platform architecture: staged grading, a provider gateway, chained investigation agents/model roles, model hierarchy/fallback, or a results resolver.

Where Operators Diverge

Routing objective differs by deployment.

APPROACH 01

Reduce expensive LLM usage by letting earlier filters or embeddings pass easy cases and sending only uncertain/suspicious cases onward.

Kalvium LabsThumbtack

APPROACH 02

Improve reliability and capacity by falling back to different models, regions, endpoints, or on-demand capacity.

Slack

APPROACH 03

Centralize access, authentication, authorization, and provider integration behind a gateway.

Grab

Model-selection strategy differs.

APPROACH 01

Fixed stage-specific model choices: Claude 3.5 Sonnet for rubric extraction, GPT-4o for scoring, and text-embedding-3-small for semantic screening.

Kalvium Labs

APPROACH 02

Cost-tiered model roles: low-, medium-, and high-cost models for expert, critic, and director functions.

Slack

APPROACH 03

User- or use-case-level model choice through a gateway exposing multiple LLM providers and APIs.

Grab

APPROACH 04

Blend rule-based checks, a legacy CNN pre-filter, and a fine-tuned LLM for the harder classification cases.

Thumbtack

Operators split on fine-tuning versus calibration/prompted orchestration.

APPROACH 01

Avoid fine-tuning and maintain a calibration loop instead.

Kalvium Labs

APPROACH 02

Fine-tune an LLM on the operator’s specific dataset.

Thumbtack

The terminal action of the cascade differs.

APPROACH 01

Send unresolved or suspicious items to human review.

Kalvium LabsThumbtack

APPROACH 02

Reroute the request to another healthy model endpoint or capacity pool.

Slack

APPROACH 03

Return provider responses through the gateway with no to minimal processing after authentication, translation, and proxying.

Grab

Watch Items

Cost, throughput, or capacity limits are a recurring driver for routing: operators limit LLM calls to uncertain cases, spill excess traffic to on-demand endpoints, or report cost/showback from gateway usage.

Human review remains part of deployed cascades where confidence or policy risk is insufficient for automatic handling.

Operators report needing safeguards for low confidence, hallucinations, variability, degraded models, or endpoint limits.

Routing across providers or infrastructure requires security, compliance, authentication, authorization, and governance work before traffic moves.

Calibration or labeled data remains operational work: deployments cited manual grading samples, validation samples, corrections datasets, or expanded training datasets.

02

Implementation Menu

CURATED DEFAULTS
NameKindMaturity
Tiered model policypatternestablished
OpenRouterserviceemerging
03

Observed in Production

2 APPS
EducationCROSS-VALIDATED

AI-Assisted Education Evaluation Review

Kalvium Labs1 OP