Microsoft Fabric accelerators, evaluation harnesses for data agents, and offline-first mobile apps.
Most of what I ship is either a one-command accelerator that deletes a day of portal clicking, or a harness that catches a regression before a user does. The rest is mobile software for languages and places that nobody writes software for.
| Project | What it does |
|---|---|
Data Agent Quality LabMicrosoft Fabric Python Jupyter Power BI |
Scheduled regression testing for Fabric data agents. Ground-truth question banks run on a timer, accuracy is trended per agent, and the pipeline fails the run when a score drops past its threshold. A Power BI scorecard then names the cause of each wrong answer: wrong query, summarization mismatch, no data returned, agent error. |
Fabric Demo GalleryMicrosoft Fabric Python Power BI Azure |
Pick an industry, sign in with Entra, click Deploy, and a full Fabric environment lands in your own tenant: workspace, lakehouse, notebooks, semantic models, Power BI reports. Twelve industries and six custom scenarios, including Real-Time Intelligence, zero-ETL mirroring from Azure SQL, and a Fabric IQ ontology behind a data agent. Progress streams live, and a failed deploy tears itself back down. fabricdemogallery.com |
FêrbûnExpo React Native TypeScript |
Free Kurdish learning app for iOS and Android. 292 Kurmancî words across 17 themes, 40 lessons, 14 interactive stories with word-level glosses, spaced repetition, streaks and badges. No account, no backend, works with the plane on. A Sorani track is being authored beside it, where every taught entry cites a reference grammar and the build rejects a citation that does not resolve. Get the app |
NisibisExpo React Native TypeScript |
Gamified city guide to Mardin and Nusaybin. Quest-driven map with live location, curated routes, per-city tour progress, and every string in Turkish, English, and Arabic. Also offline, also no backend. Site |
Microsoft data stack
Languages
Product and delivery
Field notes
Kızıltepe Ulu Camii spent a while sitting 451 metres from itself. The coordinate came from Turkish Wikipedia, and Wikipedia was wrong. The real one came off an OpenStreetMap way tagged start_date 1205. The old point sat between numbered streets in Yeni Mahalle with no building within 120 metres of it. Blast radius after the fix: 15 of 15 nearby-place lists unchanged, one route centre moved 77 metres on a 50 km route.
Two places were apologising for accuracy they already had. Mardin Kalesi and the Sakıp Sabancı museum were flagged unverified, so the app kept showing a "location approximate" badge. Their approximate values turned out to be 0 and 10 metres off OpenStreetMap.
The abbaras are still marked approximate, deliberately. They are not a building. They are vaulted passages threaded through the old city, and no single point is honest about that.
Fêrbûn's citation checker tells you what it cannot do. It proves a page reference points somewhere the book actually goes. It will not pretend to know whether the sentence on that page is correct. That still needs a speaker.
The unfinished half of Fêrbûn is unfinished on purpose. Sorani lessons nobody has written yet register as real empty lessons instead of crashing, which seemed like the more polite way for a language to be under construction.