Research and engineering projects

A susceptibility model for Rangamati Hill Tracts (a district within Chattogram division, Bangladesh), built over eight terrain and vegetation factors and served from a Flask application that assesses coordinates within the study area. It also produces a district map: 5,067 grid cells at about 1.1 km resolution.

Input mismatch

The training columns are not measurements. They are reclassified class ratings, so SLOPE is one of five integers rather than an angle. The deployed application was feeding the model raw satellite readings, which meant every live answer was computed from inputs the model had never seen, and nothing raised an error.

Implementation

Replaced a single train and test split with repeated stratified cross-validation, 5 folds by 6 repeats, and added the most-frequent-class baseline the original lacked. Selected the model with TOS rather than accuracy. Reconstructed the missing conversion from measurement to class rating, which is partly recoverable: each factor's ratings sum to exactly 100, and the rating rises with the observed landslide rate, so the class count and ordering fall out of the data even though the cut points were never recorded.

Limitation

The training inventory contains equal numbers of landslide and non-landslide sites. The output is a relative susceptibility score, not a calibrated probability that a hillside will fail. The application labels it accordingly.

Next

Recover the original reclassification table, and set the decision threshold with the same cost asymmetry the selection rule uses. Choosing the model on a false-negative-weighted criterion and then thresholding at 0.5 is inconsistent.

Repository

The landslide inventory is not mine. It is reused under licence from Rabby and Li (2020), who mapped it from satellite imagery and field survey for an area that had no official inventory.

A selection metric that scores candidates against the pool they compete in, requiring an accuracy axis and an error axis to be named separately. I was first author of the paper published at IEEE RAAICON 2021.

Why

It came out of the thesis above. Several candidate models sat within one standard deviation of each other on accuracy, and accuracy had no way to express that a missed landslide costs far more than a false alarm.

Open problem

The metric is pool-relative, so it is not independent of irrelevant alternatives: adding a candidate that cannot win may still reorder the ones above it. Demonstrable on the thesis data.

Read the full write-upTry the interactive demo DOI

A real-time dashboard over Jira's REST API tracking experiment release frequency, developer hours and team progress, used by internal teams and stakeholders to adjust sprint planning.

API constraints

Jira's API paginates and rate limits, and the interesting quantities are derived rather than stored, so the work was in aggregating issue histories into release cadence and effort within the API's request limits.

A Python integration with custom webhook endpoints giving real-time bi-directional sync of tasks, statuses and timelines between Airtable and Jira, removing manual re-entry from the experiment tracking workflow.

Sync consistency

Bi-directional sync invites echo loops, where a write on one side triggers a webhook that writes back to the origin. Making updates idempotent and origin-aware was most of the work.