Next for Seldon: from LLM calls to specialized programs. Explore our upcoming direction and help shape it.
Read the thesis →Combine AI with proven software techniques.
Make production AI faster, more efficient and more reliable.
Start with an existing LLM-powered function. Compare implementations against quality, cost and latency requirements, then explicitly activate the replacement you choose.
Tutorials
Still paying a model for the same call. Try a Seldon compiled function for better cost, quality and latency.
Workflow tutorials
Put a Seldon Function inside your application
Three code-first tutorials connect an existing AI call, the Functions dashboard, hosted execution and independent application evaluation.
Sensitive-data redaction
Build a sensitive-data redaction workflow
Optimize entity identification with Seldon while keeping span validation, pseudonymization, and leakage checks in your application.
Agent trace analytics
Extract signals for trace analytics
Optimize structured workflow-signal extraction with Seldon, then validate evidence and cluster representations in your application.
Payment-risk decisions
Optimize a payment-risk classifier
Prepare payment signals, optimize a bounded risk-classification function, and assess application-level precision, recall, F1, and fraud-dollar recall.
Get started
Sign up and pilot program
New accounts
$10
Sign up with a work email and get a workspace with $10 in credits. The $10 credit grant requires a work-email sign-up; Gmail addresses do not receive it. Existing customers can log in and keep using Functions.
Sign up →Log in →Pilot program
$3,000
Up to $3,000 in credits over 90 days for a 25-company pilot. Applications are for existing workspaces.
Sign up →Existing workspaces apply in the dashboard. New accounts sign up first.
Why Seldon
Too many production LLM calls are just expensive data work. Bring one existing LLM request. Seldon compares measured implementations, and you choose which to activate.
Cost management.
Evaluate opportunities to reduce LLM spend, with quality, coverage, and latency measured alongside cost.
Effortless productionization.
Evaluate a function before integrating its hosted alias into your application.
Reliability and auditability.
Inspect versioned execution plans and their evidence. Specialized programs can combine ordinary code with learned judgments that still need evaluation.
Execution evidence.
Inspect versioned execution plans, comparison reports, and usage for the functions you activate.
Turn one LLM function you already run into a measured, hosted alternative behind a stable alias you control.
See how it works →Drop-in router
Point your base_url at Seldon.
Route LLM calls through an OpenAI-compatible endpoint with cost, latency, and fallback controls. No SDK changes — swap the base URL and keep shipping.
Seldon base URL
SELDON_BASE_URL=https://api.seldon-ai.com/v1Use this as the base URL in any OpenAI-compatible SDK. It already includes /v1.
Explicit functions
Start with one existing LLM request.
Paste a request or short workflow into Functions. Review its inferred contract and authorize a bounded optimization run before changing your application.
Program synthesis
Evaluate a replacement before integrating it.
Functions publishes a comparison report of up to three executable alternatives. Quality, cost and latency stay visible, and you activate the one that fits. Older runs keep their original qualification rules.
FAQ
Questions from teams with real LLM traffic.
The short version: start with an existing request, read the comparison, and activate an alternative only when you choose to.
What is Seldon?
Seldon optimizes explicit LLM-powered functions and hosts qualified replacements. Its OpenAI-compatible gateway remains available; traffic discovery and new Audit analysis are retired.
How is Seldon different from a normal AI gateway?
Alongside model routing, Seldon lets you evaluate implementations for a function you identify. Explicit activation makes a replacement callable; it does not automatically intercept your agent traffic.
What's the long-term vision?
We are exploring an AI compiler and runtime around existing application functions. A typed intermediate representation would connect the intended behavior to candidate implementations using ordinary code, specialist models, and LLMs. The thesis sets out the proposal and the questions we still need to answer.
Why combine AI with proven software techniques?
An LLM call can combine semantic judgment with extraction, lookup, and arithmetic. Making those operations explicit can reveal alternative implementations. Each alternative still needs evaluation against the behavior the application requires.
How do I start using it?
If you already have an account, open Functions, save an existing request, review its preview and optimize under your chosen budget. Try a result, activate it and call its function alias. New accounts that sign up with a work email get a workspace with $10 in credits. The $10 credit grant requires a work-email sign-up; Gmail addresses do not receive it.
What happened to Live Audit?
New Audit and discovery work is retired. Historical reports and existing legacy-path controls remain available. Use Functions to optimize an operation you identify.
How does optimization actually roll out?
A finished run shows up to three measured alternatives beside your current implementation. Activating one makes that revision callable through its hosted alias. Activation does not edit your application or move traffic on its own. Evaluation on generated examples measures reference agreement or rubric pass rate, not established quality on your production traffic. Runs created before comparison reports keep their original qualification rules. Existing fallback and input limits still apply.
What data does Seldon retain?
Seldon retains API data needed for billing, reliability, debugging, audits, cache visibility, and compiled-path analysis. Trace payload capture redacts sensitive headers and secret-like values before storage.
Do you train models on customer data?
No. Seldon does not use customer prompts, outputs, traces, or cached responses to train generalized AI models.
Can we try Seldon safely in staging first?
Yes. Try an activated function with representative inputs, then integrate its alias in staging. Inspect its results and usage before changing production.