blog

Engineering notes

What I learn building agentic systems, MCP integrations and voice AI in production.

#MCP#AI Agents#Kubernetes

The agent that's not allowed to touch anything: AI root cause for Kubernetes

Why I built KubeTrace's investigation agent to be read-only by design, and why that one constraint shaped everything else.

Read

#Voice AI#AI Agents#LiveKit

Latency is the product: building a real-time AI voice interviewer

What I learned building Rubricate's voice agent, where a 600 ms pause matters more than which model you picked.

Read

#AI Evals#LLMOps#Observability

Who grades the grader? Practical LLM evals for an AI that scores humans

Rubrics, traces, human review and red-teaming: the evaluation loop I use when an LLM's output decides someone's outcome.

Read

#Vision-Language Models#AI Evals#Computer Vision

Stop picking VLMs from leaderboards: benchmark them on your own footage

Building a one-command benchmark to choose the right vision-language model for labelling video, and what generic leaderboards miss.

Read

#Embodied AI#Computer Vision#Data Pipelines

Teaching robots by watching people: an egocentric data pipeline for embodied AI

Turning head- and wrist-camera recordings plus IMU data into training datasets for robotics and world models, in fourteen swappable stages.

Read

#AI Agents#Automation#LLMs

From a list of URLs to a researched pipeline: AI agents for B2B prospecting

How I turned hours of manual account research into an automated workflow with CRM checks, people search, deep research and grounded email drafting.

Read