CONCEPT Cited by 2 sources
Model-agnostic ML platform¶
Definition¶
A model-agnostic ML platform is a shared internal service that abstracts over a specific class of models (LLMs, image generators, embedding models, speech models, rerankers) so that callers express intent ("generate a food image", "embed this document") and the platform — not the caller — owns the choice of specific model, provider, credentials, parameters, and fallback behaviour.
Two defining properties:
- Callers don't know (or care) which model they hit. Switching models is a configuration change inside the platform, not a caller-side code change.
- Platform-level investments compound across callers. Prompt-template libraries, evaluation harnesses, cost monitoring, fallback logic, rate limits, audit logs — each built once, used by every caller.
Why model-agnostic instead of model-opinionated¶
A common alternative stance is the model-opinionated platform: "we standardise on Model X; everyone uses Model X; that's the platform." This is simpler to build and reason about, but brittle in the face of:
- Different workloads have different winners. Instacart PIXEL explicitly discovered "the best performing model varied project by project." A model-opinionated platform forces the organisation to ship sub-optimal results for every workload whose winner isn't the standardised model.
- Model landscape churn. Image generation + LLMs see new dominant models every few months. A model-agnostic platform absorbs a new model behind its existing parameter protocol; a model-opinionated platform requires a migration every time the winner changes.
- Provider risk. Vendor API changes, price changes, outages, policy changes. Model-agnostic platforms can redirect to alternatives without caller visibility.
Core mechanisms¶
Model-agnostic platforms converge on a small set of mechanisms:
- Unified parameter protocol — one parameter vocabulary, platform-side translation to per-provider native APIs.
- Single endpoint — one URL / RPC service / SDK binding that fronts all models.
- Per-application defaults with overrides — working defaults ship for common cases; teams can override.
- Evaluation harness — centralised quality evaluation (human-judge sets, VLM-as-judge, LLM-as-judge) that makes model swaps quantitatively comparable.
Archetype¶
"Teams using PIXEL have witnessed a 10x reduction in the time taken to generate new imagery along with a notable increase in overall quality. […] It's easy to change to a different model and adjust prompts — so teams can move fast without needing specific model training. […] PIXEL enabled project leads to initiate projects using pre-configured, optimal model and parameter recommendations. Subsequently, they could rapidly test other models with a sample dataset and decide which one works best before moving to production image generation at scale." — Instacart PIXEL (Source: sources/2025-07-17-instacart-introducing-pixel-instacarts-unified-image-generation-platform)
Seen in¶
-
sources/2025-07-17-instacart-introducing-pixel-instacarts-unified-image-generation-platform — canonical wiki instance at the image-generation layer. PIXEL's five-component architecture (unified parameter protocol + prompt templates + DreamBooth fine-tunes + VLM evaluation + infra integration) is the archetype.
-
sources/2025-08-01-instacart-scaling-catalog-attribute-extraction-with-multi-modal-llms — second Instacart instance at the structured-extraction layer. PARSE exposes LLM choice as a per-attribute config so that simple attributes run on cheap LLMs (-70% cost at equivalent quality) and hard attributes reserve the expensive model (-60% accuracy drop on cheap LLM). Applied twice at Instacart, the model-agnostic stance is what makes per-task cost/quality tuning cheap enough to be a config field rather than a re-architecture.
Related¶
- concepts/unified-parameter-protocol — core mechanism
- concepts/cross-model-portability — immediate consequence
- concepts/single-endpoint-abstraction — core mechanism
- concepts/self-serve-generative-ai — caller-facing UX shape
- patterns/ai-gateway-provider-abstraction — text-LLM sibling pattern
- patterns/unified-image-generation-platform — image-gen sibling pattern (this wiki's archetype)
- patterns/unified-inference-binding — SDK-level sibling pattern
- patterns/centralized-embedding-platform — embedding-layer sibling pattern
- systems/instacart-pixel — canonical production instance
- systems/instacart-parse — structured-extraction sibling; per-attribute model choice as a first-class config
- companies/instacart