Pasindu Lanka · AI engineer · ColomboNotes / Lab / Shelf
I build with AI. Then I write down what broke.
Right now.
Updated Oct 4, 2026
- Building01
- Small MCP servers, to see how the protocol behaves outside the happy path.
- Reading02
- Engineering write-ups and papers on agent reliability and evals.
- Learning03
- Inference infrastructure: how requests actually get batched and scheduled.
- Wondering04
- What “done” even means for a system that is right ninety percent of the time.
Open threads.
01Agents that fail well
The question
What should an agent do when it is unsure, stuck, or wrong, and how does the system around it find out?
What I'm doing about it
Tool-call loops with explicit budgets, checkpoints and a human escape hatch. Collecting failure modes as I hit them.
Open question · Since · #agents #reliability
02MCP, from the server side
The question
Where is the Model Context Protocol genuinely elegant, and where does it leak?
What I'm doing about it
Writing small servers that expose real tools and resources, then watching how different clients treat them.
Building · Since · #mcp #protocols
03RAG, evals first
The question
Can I write the test set before the retriever, and let it decide the architecture?
What I'm doing about it
Hand-labelled questions, retrieval metrics before generation quality, chunking treated as an experiment instead of a default.
Open question · Since · #rag #evals
04The real cost of a token
The question
What does one LLM request cost end to end once caching, batching, retries and latency are all counted?
What I'm doing about it
Instrument every call, then build a back-of-envelope model I can argue with.
Open question · Since · #infra #cost
05How small is small enough?
The question
Which of my everyday tasks could a small local model do well enough that I stop calling the big one?
What I'm doing about it
Log what I actually ask for a week, then replay it against smaller models. Waiting on free time.
Parked · Since · #local #models
Everything, newest first.
- 27Sept ’26The real cost of a tokenWhat does one LLM request cost end to end once caching, batching, retries and latency are all counted?LabOpen question
- 21Sept ’26Model Context ProtocolRead the spec itself, not only the tutorials. It is shorter than you think.Foundmodelcontextprotocol.io
- 20Sept ’26MCP, from the server sideWhere is the Model Context Protocol genuinely elegant, and where does it leak?LabBuilding
- 18Sept ’26Building effective agentsThe clearest argument for starting simple: workflows first, agents only when you must.FoundAnthropic
- 12Sept ’26Agents that fail wellWhat should an agent do when it is unsure, stuck, or wrong, and how does the system around it find out?LabOpen question
- 2Sept ’26Your AI product needs evalsWhy the unglamorous part, looking at your data, is the whole job.FoundHamel Husain
- 30Aug ’26RAG, evals firstCan I write the test set before the retriever, and let it decide the architecture?LabOpen question
- 26Aug ’26Hello, worldWhy lanka.lol exists, and what it is deliberately not.Note1 min
- 26Aug ’26How this site is builtNext.js, Tailwind, markdown files and a handful of small client components. Notes on a site that tries to stay out of the way.Note1 min
- 20Aug ’26The Bitter LessonTwo pages I reread whenever I am tempted to be clever.FoundRich Sutton
- 18Aug ’26Attention Is All You NeedStill worth reading in the original, if only to see how plain the idea looks on paper.FoundVaswani et al.
- 16Aug ’26Designing Data-Intensive ApplicationsThe best map of the ground underneath every AI system I have worked on.FoundMartin Kleppmann
- 14Aug ’26How small is small enough?Which of my everyday tasks could a small local model do well enough that I stop calling the big one?LabParked