github.com/exoxeph
Dhaka, Bangladesh
exoxephImtiaz Mashrafee
Each repository here has a case study: what was built, what I contributed, what was measured and what was not.
- Handle
- exoxeph
- Role
- AI Engineer
- Studying
- Computer Science, BRAC University
- Case studies
- 4, each with evidence states
- Also public
- 2 smaller repositories
- Languages
- PythonTypeScript
- Verified
- Partially verified
- Not claimed
Case studies
exoxeph /email-assistantPublic
Two-stage email drafting pipeline
An AI email writer that first checks whether the request makes sense, and asks a question instead of writing when it does not.
Checked by me Ten test scenarios are saved in the repository, with the recorded outputs and the scores a judge model gave them.
Case study: Two-stage email drafting pipelineRepository: Two-stage email drafting pipeline
exoxeph /Data-Aware-RAGPublic
Retrieval over research papers with verify and repair
An AI that answers questions about research papers, scores its own draft for quality, and gets one retry when the draft falls short.
Checked by me Saved evaluation runs record which path each question took through the pipeline.
Case study: Retrieval over research papers with verify and repairRepository: Retrieval over research papers with verify and repair
exoxeph /speech-doc-extractionPublic
Speech and report extraction that does not invent values
A system that transcribes speech and pulls structured data out of photographed lab reports, leaving out any value it cannot read instead of guessing.
Checked by me Test recordings and synthetic report images are saved in the repository.
Case study: Speech and report extraction that does not invent valuesRepository: Speech and report extraction that does not invent values
exoxeph /tesla-poolPublic
Ride-pooling lifecycle with guarded transitions
A ride-sharing backend that refuses a booking when the car would end up with more riders than seats. It does not yet hold when two bookings arrive at the same instant.
Checked by me 71 written test cases exist in the repository. They were counted, not run for this page.
Case study: Ride-pooling lifecycle with guarded transitionsInspect the system (Ride-pooling lifecycle with guarded transitions)Repository: Ride-pooling lifecycle with guarded transitions
Also on GitHub
exoxeph /email-spam-classificationPublic
A four-version spam classifier on the Enron dataset, from a TF-IDF and Naive Bayes baseline to a threshold-tuned comparison. I document a test-set leak in an early version and correct it with a validation split (self-reported F1 0.9894, test-selected, then 0.9891, validation-tuned; not independently reproduced).
exoxeph /CryptonitePublic
An educational course cryptography demonstrator with toy parameters and documented AI-assisted planning. The project states it is unsuitable for real data.
Not claimed These projects have no case study on this site.