I build AI platforms and agent infrastructure.

• AI in simple terms: creating a basic agent harness

  14 Sept 2026·4 min read

  Hi Guys, as promised in the previous articles, I'm going to show you how to create a dead simple agent harness in this tutorial…

• AI in simple terms: what is an MCP server?

  5 Sept 2026·4 min read

  Hi Guys, in the last article I described what an agent harness is, and I promised we would build one next. Then a friend told me MCPs are dead and agents just use CLI commands now. I didn't want to argue — so I wrote this instead…

• AI in simple terms: what is an agent harness?

  27 Aug 2026·4 min read

  Hi Guys, in this article I'm going to simply describe what an AI harness is. Everybody is so overwhelmed with so many buzzwords, so this one is to make it easy for anyone using a coding agent to understand what they are really using…

• Your blog doesn't need a vector database

  20 Aug 2026·8 min read

  I wanted to ask my own archive questions. The obvious 2026 answer is embeddings, a vector store and an LLM on top — so I measured the corpus first, and the numbers made all three look ridiculous…

• marsClaw — Building a personal chat assistant with Bun, SQLite and MCP

  18 Aug 2026·11 min read

  Hey Guys, for the past few months I have been building a small thing called marsClaw. It is a personal assistant which lives inside a chat app…

• Why MCP, Not REST APIs?

  3 Aug 2026·10 min read

  The question comes up the moment you try to make an agent use your existing API. The answer isn't that REST is wrong — it's that REST was designed for a client who reads the docs at build time, and an LLM isn't that client.

Things I've built

• platform-mcpRead-only MCP server exposing 15 GCP observability tools across logs, metrics, cost and asset inventory, so an LLM agent can triage incidents and find savings without ever mutating infrastructure. In production use by my team.
• bigquery-mcpMCP server for plain-language questions over BigQuery. Dry-run cost estimation, tiered confirmation, a hard per-query scan cap, and SELECT-only execution under each user’s own IAM credentials.
• LimbiqNeurotransmitter-inspired memory layer that makes any LLM appear to learn across conversations — FAISS vector search plus a knowledge graph, with no change to model weights. Published on PyPI.
• marsClawMulti-channel LLM agent across Telegram, Slack and WhatsApp on the Claude Agent SDK, with provider failover, capability-removal-by-default security, and an append-only tool audit log.
• FleetOS17-service Go platform streaming 20-DOF robot telemetry over gRPC into Kafka, with a Temporal-orchestrated ML lifecycle running from experience collection through GPU training to canary rollout with automatic rollback.
• kube-gpukubectl plugin giving real-time visibility into NVIDIA MIG partitions — per-pod GPU-slice assignment, pending job queues, and CUDA/NCCL/OOM highlighting in streamed logs. Ships via Homebrew and krew.
• KithlyConsumer product built end-to-end and solo: a React Native app, a Go API on Clean Architecture, pre-signed S3 uploads, and a Redis/asynq worker pipeline that stitches contributor clips into one finished video.
• ZylobinCivic recycling platform built solo across four codebases — React Native app, admin console, marketing site and Node API — covering 1,500+ verified collection bins across a 15-module domain.