ARGet the notes
AR

Hi, I'm Ayesh.

I build AI search for products that already have users.

Search that understands what your customers mean, over your own data, plus the backend it needs to hold up.

If you like real numbers, honest write-ups and the parts that failed, you'll feel at home here.

This site is where I show my own projects in full and write one engineering note a week.

Backend. Search. Numbers.

Hey, thanks for reading this far.

I'm Ayesh Rehman, a backend and AI search engineer in Karachi. I've spent more than seven years building backends and the apps on top of them, from news apps with millions of installs to the AI search I work on today.

In my day job I build production AI search for a US company, under a non-disclosure agreement. So everything here comes from my own projects, which I can show in full: the code, the numbers and what failed.

At the core of how I work is one rule:

Every change to a search system gets a number before and a number after.

I write it all down, one note a week, because learning in public is the most useful thing I can give back. More about me.

AR Ayesh Rehman

years building software, since January 2019
7+
installs across the Geo News, Geo Super and Jang apps I delivered and maintained at Jang Media
20M+
apps moved onto one sign-in, in an SSO migration I led
5
engineers mentored across two teams
5
Projects

From a problem I had to shipped code

Some are tools I use every day. Some are experiments I measured. All of them are mine to show.

hit@10, before and after enrichment
Question setBeforeAfter
Main question set0.700.80
Held-out questions0.450.75
Private repo · evaluation report coming

Nooor

Retrieval I can put a number on.

Retrieval over Quran tafsir, measured on fixed question sets. Enriching each passage before embedding raised hit@10 on both sets.

Embeddings · retrieval evaluation

The Code Akos dashboard: 9 errors tracked, 9 triaged, 9 filed on GitHub, 9 fixes tried and 4 pull requests open.Source available · v1.3.0

Code Akos

Errors in, reviewed pull requests out.

Self-hosted, Sentry-compatible error tracking. It groups the errors your app sends, files a GitHub issue for each new one, and has an LLM attempt a fix that it opens as a pull request for you to review. It runs your check command before and after the fix and marks the pull request verified or unverified.

Node.js · SQLite · Ollama, DeepSeek or Anthropic

Explore the code Watch the demo

Source available · v0.1.2

Muniko Share

Phone to computer, with no server in between.

Files and clipboard between a computer and a phone on the same network. Devices find each other over mDNS and pair over TLS, with no server in between.

Rust · Tauri · axum · rustls

Explore the code

Live tracker

Always building something

Evenings and weekends included.

GitHub

693contributions on GitHub in the last year
Writing

Somewhere between a build log and a lab report

I write to check my own work and pass on what I learned. These are the first four notes.

  1. Coming soonHow I measure whether a RAG system is getting better
  2. Coming soonWhy my held-out test set showed a bigger gain than my main set, and why that worried me
  3. Coming soonTurning production errors into pull requests: the Code Akos architecture
  4. Coming soonWhat I let the model change and what I never let it change
  • Measuring RAG quality
  • Search over your own data
  • Held-out test sets
  • Errors into pull requests
  • Guardrails for model fixes
  • Reviewing AI pull requests
  • Boring backends
The notes

One note a week. Real numbers.

Each note takes one problem from my own projects and shows how I solved it, with the scores and the parts that failed.

One note a week on AI search. Unsubscribe from any email.

What I build with

No sponsorships and no referral links. Each tool here is one I've shipped with, and the card says where, so you can check.

Filters

Showing 10 of 10

  • Backend

    FastAPI

    AI search APIs, from the resume

  • Backend

    Django

    Production backends, from the resume

  • Backend

    Node.js

    Code Akos server

  • Data

    SQLite

    Code Akos storage

  • AI

    Azure OpenAI

    Production AI search, from the resume

  • AI

    Embeddings

    Nooor retrieval

  • AI

    Ollama

    Code Akos local models

  • Apps

    Rust

    Muniko Share

  • Apps

    Tauri

    Muniko Share desktop app

  • Apps

    React Native

    Geo News and Jang apps, from the resume