Fifteen years in production software, ten of them at Innovaccer on clinical data platforms. Now independent, consulting on data and AI, and building in the open.
Most of what I build turns out to be the same problem in a new language: move something through a pipeline you can still inspect afterwards. Started with messages off a Redis queue in 2017, ended up with a grounding layer that has to justify why it cleared a finding.
Lithrim CE (Apache-2.0) Evaluation harness for agentic AI. A configurable LLM judge council sitting over a deterministic grounding layer, where a flag stands unless a machine-checkable authority positively disproves it, and an unresolved check never clears anything. Every grade writes an immutable record with the configuration hashed into a signature, so a past result reproduces from storage with no model call. Study on Zenodo, preregistered on OSF. The rerun rejected three of my four predictions and I published that as a new version.
ETLP Clojure data-pipeline library built on transducers and core.async. Protocol-based connector contract, DAG executor, workflow DSL, and a streaming HL7v2 parser. Written as a side project in 2020, then it turned out to solve a production ingestion problem and went to work.
Zyng Turns a product into an editable walkthrough or launch film. Live and self-serve.
Mapify was an agentic UI over low-code data transformation, an attempt to package ETLP so a data engineer could drive it without writing the transform by hand. It is not running. Part of that work became Lithrim's ingestion layer.
I build things fast and I am not precious about killing them. The parts tend to come back.
- Breaking Down the Last Monolith: Micro Frontends.
- Lamdba Calculus
- Technical reviewer on The Art of MicroFrontends, first edition.