Level 1 of 7
Apps
The apps people use on their phones, both my own and for clients.
RockLens
Take a photo of a rock, crystal or gem and get the three most likely matches, each labelled by how confident the app is. Live on the App Store with paying subscribers.5
Countstead
Stock tracking that works with no internet connection. It keeps a permanent record of every change and comes in 9 languages. Live on the App Store with paying subscribers.5
Imprimo
Exam flashcards made with AI.
Language-learning app
I'm the lead mobile developer. It has 50k+ users and earns five-figure monthly recurring revenue in US dollars.2
Stock management and ordering system
For a multinational company in the Nordic region.6
EatMealFirst
Food delivery with separate apps for customers, restaurants and riders. It handled 800+ orders a week in Sri Lanka and Canada. I also led a team of 5 developers for a few months.7
Taxi-hailing apps
For a major financial services provider in Sri Lanka, with live tracking and billing.
Fleet management app
Live vehicle tracking, vehicle history and reminders.
ShiftSum
Privately tracks your hours and estimates your pay before deductions.
Level 2 of 7
Servers and connections
The systems behind an app. They store data and connect it to other services.
RockLens server
Sits between the app and the AI. It limits how much each device can use, checks that requests come from the real app, and has a spending cutoff.
Countstead server
Stores data and handles subscription updates.
Imprimo server
Runs the app's online features, with background jobs and an admin dashboard.
- iPhone appProves it's the real app
- RockLens serverApplies the limits and the spending cutoff
- GeminiGoogle's AI model
Level 3 of 7
Automation
Software that handles repetitive work so nobody has to do it by hand.
Odoo Teams bot
Type your hours in plain English in Microsoft Teams, and the bot enters them in Odoo, a business management system.
Daily social media pipeline
Every day it checks RockLens's posting queues, creates and checks new videos, and tops up Instagram, YouTube, Pinterest and Facebook. A separate job checks that posts went out and recovers any that failed.
- Check the queues
- Create and check videos
- Top up the channelsInstagram, YouTube, Pinterest, Facebook
- Separate jobChecks posts and recovers failures
Level 4 of 7
AI features
Where AI does a useful job inside a product.
RockLens photo identification
The AI looks at a photo and returns its three best matches, ranked.
Imprimo flashcards
AI writes the flashcards, and a method called spaced repetition schedules when you review each one.
Reading plain English in the Teams bot
The bot narrows thousands of tasks down to the 5 closest matches, so the AI only has to choose from those.
- Your message
- Thousands of tasksNarrowed down
- Top 5 matches
- Gemini
- Timesheet entry in Odoo
Level 5 of 7
Testing the AI
How I check that an AI is good enough before anyone relies on it.
RockLens release check
A new version only ships if the right answer is in the top 3 at least 85% of the time and unusable photos are caught at least 95% of the time, checked across 12 languages.8
model-evals
My own test suite for AI models that run on your own computer. Each model has to complete tasks by itself, and a program scores the results, with no AI doing the marking. The full version has 183 tasks.9
- Right answer in the top 3At least 85%
- Unusable photos caughtAt least 95%
- Checked across 12 languages
Level 6 of 7
Running the AI
The engine that runs an AI model on a computer.
My version of Strata
Strata is an engine that runs a very large AI model on a gaming PC. It's by Niko1221, and I made my own version tuned for the RX 7900 XTX graphics card.10 Tested against the original v0.2.4.1 on the same PC on the same day, in alternating rounds, median of 3. An optional build setting was on for this test.4
Measure Original My version Change Writing speed, short chat 82.6 tok/s 97.5 tok/s +18% Time to start replying, short chat 1.20 s 0.84 s -30% Reading a long document (32K tokens) 1,359 tok/s 1,928 tok/s +42% Time to start replying, AI agent session 3.98 s 2.99 s -25% A smaller optional format
With part of the model stored in a smaller format called EXL3, answers are 32% faster, replies start 66% sooner and long documents are read 63% faster. That part of the model takes 22% less space, 39.5 GB instead of 50.3 GB.11
- Reading a long document1,359 to 1,928 tok/s, +42%
- Starting to reply1.20 s to 0.84 s, -30%
- Writing the answer82.6 to 97.5 tok/s, +18%
Level 7 of 7
The chip
The lowest level: small programs, called kernels, that run directly on the graphics card's chip.
Hand-tuned kernels
I hand-tuned these programs for the RX 7900 XTX.
Faster, with the same output
These changes made chat answers 11.7% faster with the same output, tested at 2,557 positions. I proposed them to the original project as PR #2231, and it was closed without merging.12
How I got here
Made games in Unity at Crowderia AB, including Roofdown, a platformer for a Swedish rock band, and Twingo, puzzle mini-games set in London.
Built Android apps at IDAMA Innovations: four apps including a service marketplace, sales-team location tracking and a point-of-sale system.
Built a food delivery system at GlobalValley Solutions that fast food outlets in Mexico could brand as their own.
Senior Android engineer at ThinkSmart Solutions, building EatMealFirst and leading a team of 5 for a few months.
Senior Android engineer at Nimbus Venture, working on taxi-hailing and fleet management apps and a grocery ordering proof of concept in Flutter.
Technical Lead - Mobile at Allion Technologies.
Shipped my own iPhone apps and tuned an AI engine down to the chip.
Contact
- Emailimdad@spacewa.lk
- LinkedInlinkedin.com/in/imdad-ismail
- GitHubgithub.com/android-imdad
- App StoreMy developer page