Blog
Notes from the router and the compiler.
How Seldon routes, observes, and progressively compiles repeated LLM workflows into cheaper, deterministic data pipelines.
The state of decision models, Part 1: adoption, competition, and the integration gap
Decision models are spreading fast. Detailed case studies reveal where they save work, where replacements fail, and the engineering patterns behind dependable adoption.
Before you fine-tune, compile: lessons from a payment-risk experiment
A payment-risk function achieved higher F1 at roughly 27× lower inference cost than Luna, without custom weight training. What worked, what failed, and why Seldon optimizes implementations around business targets.
We fired the frontier model from two AI pipelines. Here are the receipts.
Seldon Functions compiled the LLM call at the center of two hard data pipelines. Agent-trace extraction got 5× cheaper and more accurate; PII detection got ~8,000× cheaper and up to 10× faster. Here are the production receipts, including what broke.
Semantic Compute: From Interpreters to Compilers
Jev's traction points toward a broader programming paradigm: semantic functions, a typed intermediate representation, and compilers that choose among models and ordinary software.
AI applications need better primitives than text generation
A production LLM call is often a latent program. Name the operation. Search the traffic for that program. Compile each step. Leave the frontier model for what is new.
The great unbundling of the LLM
A model that refuses to write a single word became the most-copied idea in AI in under 48 hours. Everyone called it a cheap classifier. They missed the story. Jev is the first crack in a monolith that was always going to come apart — and the fragmentation it kicked off is the one we bet the company on.
Observational reconstruction, then the inverse problem
Why we capture LLM traces, why reconstructing a session is not enough to compile a step, the technical challenges that follow, and what Seldon has actually shipped and measured so far.
The AI bill is becoming a management discipline
Tokenomics and FinOps make AI spend visible — but the harder problem is deciding which workflows should remain model calls at all. Why counting tokens is only the beginning.
You don't need an LLM to cluster LLM traces
How Seldon's Trace Audit finds compilable workloads with contract-first density clustering — and why summarize-then-embed made the clusters worse on a 12.5k-trace ablation.
How much of your LLM bill is just ETL?
A technical essay on measuring production LLM traffic: how to distinguish reasoning from routing, extraction, classification, normalization, lookup, and deterministic data work.
The Silent Epidemic of LLM Technical Debt
Prompt-driven development quietly buries core application logic inside probabilistic text. Here is how to tell which LLM calls have matured into stable workflows that should become deterministic systems.
Program synthesis, and why compiling LLM calls into ETL is one
What program synthesis is, and how Seldon's approach of turning repeated LLM calls into deterministic data pipelines is a restricted, example-driven form of it.