Dendrite

Dendrite is a research platform for testing language models. It supports doing things like concept injection for the SPAR research and it also supports neural architecture search for RNN's. It is designed to allow me to queue up multiple research and training jobs which will keep my GPU cluster busy.

For the research side of things an experiment is a single config file that describes a grid of conditions. Dendrite turns each point of the grid into trials, schedules them onto GPU hosts, and keeps every result in Postgres as the one source of truth. Further information is stored in Garage a S3 compatible storage service.

Dendrite keeps the scheduler but it launches runs with Argo Workflows. Argo Workflows launches a Python worker on a GPU host that loads the model, performs the injection, captures activations and generates text. Dendrite's control plane and the browser interface are TypeScript with Effect, and the whole thing deploys to my homelab cluster with Pulumi. Dendrite also has built in LLM-as-Judge for responses. Initial calibration of the judge is done by loading a dataset onto my dataset-review service were label a sample. The LLM-as-Judge is then prompt engineered by another agent to match my hand labeling.

I built the MVP in about a week in August 2026 for a research interview demo, and the MVP went live in the cluster on August 28. Before continuing, I compared it with UK AISI's Inspect. I kept building Dendrite because Inspect has no distributed workers, job queue, relational results store, blinded judging or injection support.

While doing the SPAR research fellowship I have been learning a lot of things that I was missing in my first attempt. Unfortunately the speed of the fellowship has prevented me from reconciling the data formats so that Dendrite can run that research. I need a complete redesign of the experiment config to handle more bespoke runs. I will be pulling in some ideas from UK AISI's Inspect with regards to running code rather than configs.

For the training side of Dendrite it is similar to KubeFlow. Unfortunately Kubeflow doesn't support training Recurrent Neural Networks or combining things with concept injection experiments. So I started development back in July 2026 however the RNN side of Dendrite has been put on hold while I focus on the SPAR Fellowship.