I build the tools that connect systems which were never designed to talk to each other.
Search for blog posts, studies, projects, and pages
Projects
NarsilNarsil is a search engine for full-text, vector, hybrid, and geospatial queries on one typed index, which you can embed in your app or deploy behind its own server.
SirannonTurn any SQLite database into a distributed data layer with live queries, offline-first device sync, multi-node replication, and coordinator-backed failover.Topics
Try retrieval, retention, languages, or seek. Narsil searches every post, project, and documentation page here and returns matches as you type.
4,200/sec
Narsil processes keyword queries at that rate on BEIR SciFact, against 2,796 for Elasticsearch. It also ranks the results more accurately than Elasticsearch and OpenSearch on the same benchmark.
5runtimes
The same replicated SQLite layer runs on better-sqlite3, the SQLite built into Node.js, wa-sqlite in the browser, Bun, and Expo. A phone keeps working offline and settles its conflicts with the server on the next sync.
20s → instant
Health coaches waited twenty seconds for each page to load. The page now opens as they ask for it, so the conversation with a young person continues.
40%lower
Redesigning the architecture cut the infrastructure bill by that much.
500+students
The COVID-19 pandemic closed the campus. Its student president kept examinations and teaching running for that many students.
Delali Azumah
What a product can do depends on the layer underneath it.
Understand one subject well enough to fix it yourself, one article a week.
What a business earns depends far less on the hours that its owner works than on the kind of work that she spends them on. Across the five articles in this study, I set out the four levels of value, from doing every job yourself to keeping only the decisions.
All 5 articles published
Your agent can still provide wrong answers after you swap the model over, rewrite the prompt, and drop the temperature. The fault is one step earlier, in the search step that returns the passages from which your agent generates its answers. By the end of this study, you can measure that step, find what limits it, and prove that the fault is gone after your change.
All 5 articles published
These articles cover engineering, business, community, and faith.
When an AI feature returns wrong or invented answers in production, the cause is usually the search step before the model. In this article, I will show you how you can measure the quality of that search and find the real cause.
10 min read
Read the articleYou can buy hands, managers, and even the selling, once you write your knowledge down. The one thing that you cannot buy is somebody to decide what the business is, so in this closing article of the study I explain why the owner has to keep that job.
Selling, when you do it honestly, is explaining your work to the person who is yet to buy it. You learn the subject by doing that work for years, so in this article I treat teaching and selling as one job.
With embeddings, a search engine can return answers that share no words with the question. In one recorded run over two public corpora, vector search on its own scores below keyword search on both corpora, while the fused ranking of the two scores above both. In this article, I will show you how you can test your own corpus to find out which result you get.
When you hire somebody, you buy back hours of your week, but whether the hire earns back their wage depends on what you refill those hours with. In this article I cover what to hand over, what you have to put in a written standard, and what you will then have to do as the manager of the work.
The rule is on page 31 of your handbook, but your agent provides a different answer. Your search engine returns whole indexed pieces of your documents, and the right size for those pieces depends on how long your documents are. In this article, I will show you how you can tell whether to cut your documents, where to cut them, and how you can prove the choice on your own questions.
Delali designed and built each of these.
Narsil is a search engine for full-text, vector, hybrid, and geospatial queries on one typed index, which you can embed in your app or deploy behind its own server.
Turn any SQLite database into a distributed data layer with live queries, offline-first device sync, multi-node replication, and coordinator-backed failover.