Independent AI research & development

Helping small models
do more reliable work.

A model shouldn’t have to improvise every step. We’re building reusable operations that put predictable work into reliable processes, so models in the 3–32B range can focus on the judgments that need them.

Request more information Explore the approach

We’re actively developing the specification and runtime and welcome private technical discussions.

01 / PODL

A shared language for
tools, models, and decisions.

PODL is a document specification for defining whole operations and the tools they use. It builds on the Model Context Protocol (MCP), adding the contracts needed to coordinate work across multiple steps. Its design targets models in the 3–32B range: capable enough to specialize, small enough to run where the work is. Define the process once; a runtime carries it out.

Remove avoidable mistakes.

Give models valid choices and let predictable steps run without another AI decision.

How constraints help

PODL lets tool authors define typed references to values and operation authors specify where those references can be used. A model can select a result instead of copying a long identifier or constructing a value itself.

Our reference runtime enforces those choices during generation and resolves the references for the tools. That removes specific ways to make a mistake before they reach the next step.

Use the right model for each job.

Describe specialist roles in the operation so a deployment can choose models that fit its environment and cost constraints.

Specialist routing in the specification

PODL supports advisory specialty hints for operations and tools. A runtime can use them to route planning, retrieval, vision, coding, or review work to different models, or ignore them and choose its own strategy.

Our reference runtime does not yet implement specialist routing. The specification keeps the choice portable while leaving model selection to the implementation.

Keep a path to the evidence.

Let tools identify the sources behind their results, so answers can be traced and checked.

From specification to implementation

PODL defines source handles and lets tools describe the provenance of their results. Our reference implementation uses those contracts to retain originals, constrain citation choices, and support review.

How that evidence is exposed for inspection is an implementation choice. A real citation can still be used incorrectly; traceability makes that error easier to investigate.

Portable Operation Definition Language

Specification and reference runtime in active development. Our implementation runs existing small models locally or through privately hosted inference endpoints.

The design principle

Give AI less
to get wrong.

Reliable processes reduce opportunities for human error. AI needs the same discipline: when a step can be computed, checked, or carried forward exactly, it shouldn’t depend on a model getting it right.

PODL lets developers put that discipline into a reusable operation. The aim is to make bad states unrepresentable wherever the process allows it, leaving the model to do the work that requires judgment.

Our longer-term goal is an operation that improves other operations: learning from good and bad outcomes to change the process so an observed failure can no longer occur. This is a research direction, still to be built.

How operations could improve

02 / Model research

How much more can
a small model do?

Alongside PODL, we are exploring and documenting how small language models learn and use computation. The aim is to find architectures worth scaling by measuring how design and training choices affect quality and computational cost.

Experimental research

Can repeated computation improve a model’s answers?

We study models that revisit an internal computation, allowing additional processing without adding a new set of model weights for each step.

The question is whether that extra work improves quality enough to justify its cost. We’re testing training choices, model size, and how learned components transfer.

Where the research stands

Our experiments focus on looping transformer models: a shared set of layers runs repeatedly, refining an internal state with each pass. We’re documenting how training and inference loop counts affect quality and compute cost, with the longer-term aim of creating smaller models focused on common specialist roles. Findings are still being tested at small scale; they don’t yet establish a production advantage.

This research is separate from our PODL reference runtime, which uses existing models.

Let’s talk

Bring us a problem
you’re working on.

Tell us what you need from local AI, evidence-backed research, or small models. We can discuss fit, current capabilities, and what remains to be built.

Architecture, evaluation results, and research details are available for discussion under NDA.

Helping small modelsdo more reliable work.

A shared language fortools, models, and decisions.