Hey, I'm Kar, glad to see you here. I'm a staff engineer who likes building systems that work as expected and outlast my utility: the kind you can hand to someone else and walk away from. I've spent 14+ years doing that across medical imaging, gas line construction inventory, financial applications for federal institutions such as Human Services and the USDA, fitness, and now education tech, and for the past eight years I've led teams of 4 to 30 on AWS and GCP.
AI-assisted development accelerated the pace of systems change, and how easily their behaviour drifts from baseline expectations. So I start from the structure rather than the output: guardrails and rules engines that bound what an agent can see and do, evals that pin down the behaviour you expect, and observability for when it diverges. GigAI is where I work this out in the open: every model call leaves a journal, and no claim ships without a source.
I build with AI every day: my own agent runtime, and a way of running agents I call orchestrator–coordinator. I trust a model to pull data out of systems and correlate it. When it presents facts and claims, I verify, because a model is a nondeterministic algorithm: the same input can come back different on the next run.
Currently
I'm a Staff Engineer at Guild Education, where I have led teams across identity, lead management, enrollments and transactions.
I built an agent-native ops platform for production support. For a single student, an agent gathers data from twelve services and from warehouse reports in Snowflake and Databricks, then correlates it against a deterministic rules engine. Every response passes through a PII redaction layer before it reaches the agent's context. The agent never mutates anything: it returns an investigation report and a proposed resolution with the exact payload to apply, such as events to resend, and a person runs it.
I was also the staff engineer overseeing the migration off a legacy monolith that had a claims engine and eight other domains built into it. It took about six quarters across multiple teams and services. I worked across teams and pillars to roll it out gradually, one learning partner at a time, until the monolith was retired.
Projects
- GigAI: an open-source, local agent runtime where every model call leaves a journal and no claim ships without a source.
PythonReactTypstCodex / Claude CLI - Scout and the learning center: GigAI's first Gig, a job search that runs on your own machine, plus a public, account-free learning center.
PythonGigAIstatic site - EdgeLab: an autonomous trading-research platform with typed research agents, an evaluation suite and cited retrieval.
GoPythonReactKubernetesTerraformPostgreSQLpgvectorRabbitMQ - Forge: a closed-loop research substrate where forecasts are committed before outcomes are known.
PythonPostgreSQLKubernetesTerraform
Blog
- The orchestrator–coordinator pattern: how I run agentic development with an orchestrator, coordinators and independent readers on a different model family.