software engineer backend · full-stack · agents Berlin

Sanath Samarasinghe - Software Engineer

sanath@dev:~
A garden, not a blog

This site is a digital garden, a set of notes on engineering, AI and the things I'm building, that link to each other and grow over time. Wander with the file tree on the left, or hit ⌘K to jump anywhere.

Recent writing

Mar 09 26
The Simple Loop Behind Every Fancy Agent
Every sophisticated agent harness still boils down to this tiny loop: call the model, run tools, feed results, repeat.
#ai#development#tools
Feb 13 26
The Three Layers Every Production RAG System Needs FEATURED
Document parsing, hybrid search, and orchestration - the three layers that separate toy RAG demos from production systems.
#ai#rag#langchain#opensource
Feb 01 26
How OpenClaw Gives AI Agents Persistent Memory With Plain Markdown
A deep dive into how OpenClaw's file-based memory system gives AI assistants genuine continuity between sessions using Markdown files, semantic search, and automatic memory flush.
#ai#openclaw#development#tools
Oct 29 25
The 5 MCP Servers Every Developer Needs in 2025 FEATURED
The 5 MCP servers that actually matter for your AI-powered dev workflow - semantic code nav, GitHub automation, browser testing, live docs, and persistent memory.
#mcp#ai#development#productivity

Now

Right now I'm building Draph and an agent runtime, writing about evals for agentic systems, and going deep on small language models. My About note has the full picture - current focus, the stack I reach for, and work history.

  • Building - Draph - diagram & flowchart tool with AI-assisted layout. Meso - an agent runtime.
  • Writing - A long post on evals for agentic systems - why standard approaches fall short when tools have side effects.
  • Learning - How to build my own SLM - going deep on the fundamentals of small language models.
  • Location - Berlin, DE.

Open source

esc