https://shovon.bd/blog Articles on machine learning, GIS and web mapping, and Android apps by Md Minaruzzaman Shovon. en-us Thu, 01 Oct 2026 00:00:00 GMT https://shovon.bd/blog/tutor-attendance-app https://shovon.bd/blog/tutor-attendance-app Thu, 01 Oct 2026 00:00:00 GMT Md Minaruzzaman Shovon Productivity Why attendance tracking gets messy for home tutors and private teachers, what a good tutor attendance app should include, and how Tuition Tracker: Attendance handles monthly class targets, missed classes, and history.

• Private tutors usually track attendance alone, across notebooks, spreadsheets, calendars and memory, so monthly class counts drift.
• The core questions each month are: how many classes are done, how many are left, which were missed, and what happened in past months.
• A good tutor attendance app needs per-student monthly targets, weekly schedules, one-tap attendance, a monthly calendar, history, and backup across devices.
• Tuition Tracker: Attendance is an Android app built for this workflow, with Days Done / Days Left counts, Month-End History, WhatsApp Quick Connect and Google cloud sync.

]]> https://shovon.bd/blog/typic-open-decision-model https://shovon.bd/blog/typic-open-decision-model Thu, 24 Sep 2026 00:00:00 GMT Md Minaruzzaman Shovon Machine Learning How I built Typic, an open 400M decision model that answers typed questions in one forward pass. Trained on Kaggle's free T4 in 46 minutes, it scored 69.9% vs Laya's 56.7% on unseen tasks and stayed accurate on 7,000-token inputs.

• Typic is an open decision model: given a context, a question and a list of options, it returns a calibrated probability for each option in a single forward pass, so there is no text to parse.
• It is built on the Ettin-400M encoder and trained on 50,000 heavily augmented examples from 25 public datasets, in 46 minutes on one free Kaggle T4 GPU.
• On 2,000 questions from four unseen tasks it scored 69.9% overall vs 56.7% for Laya, with 29 ms vs 37 ms latency.
• With up to 7,000 tokens of unrelated text before a request, Typic kept 17–18 of 20 routing answers correct while Laya dropped to 5–6.
• Limits: the author chose the tests, it is a single training run, it is English-only, and it needs temperature scaling to match Laya's calibration.

]]> https://shovon.bd/blog/chittagong-commute-time-map https://shovon.bd/blog/chittagong-commute-time-map Fri, 08 May 2026 00:00:00 GMT Md Minaruzzaman Shovon Web Development A step-by-step walkthrough of building an interactive isochrone map of Chittagong: click anywhere and see where you can drive in 10 to 60 minutes, with Mapbox's Isochrone API, Mapbox GL JS road colouring and Turf.js hover tooltips.

• The app shows every place reachable by car within 10–60 minutes from any point in Chittagong; clicking or dragging the pin recalculates instantly.
• Drive-time polygons come from the Mapbox Isochrone API, fetched client-side as two parallel requests because the API allows at most 4 contours per call.
• A GPU-evaluated Mapbox style expression colours bands from yellow (near) to navy (an hour away), and a Roads mode uses the `within` expression to paint travel time onto streets.
• Turf.js booleanPointInPolygon powers instant hover tooltips showing the road name and its drive time from the pin.
• There is no backend: it deploys statically, using a URL-restricted public Mapbox token.

]]>