AI / Agent Engineer · agent systems, and the evals that tell you when they break
🧩 Currently puzzling over
- Touchstone — an eval-improvement loop for agentic systems. Failing runs are mined back into the eval set, gates guard both directions of error, and the specimen benchmark grades final database state alone.
- Axiom — turns scraped financial data into a queryable company graph. Airflow DAGs ingest and enrich it, DSPy extraction builds the nodes and edges, and a validation DAG checks what lands.
- Autopulse — P2P lending automation: alerts, auto-invest and analytics, running as systemd units on a Terraform-provisioned GCP VM.




