Muhammad Shahwaiz
  • Research
  • Prototype

Menzync

Emotion research and student check-ins.

Menzync explores affordable, private wellbeing support for university students in Pakistan. It runs as two separate tracks: emotion-recognition research, and a student check-in prototype.

Scope & stage

Research and a local prototype, kept deliberately separate. The prototype doesn't run any emotion model yet. It is an FYP and early-stage startup project, a finalist in National Idea Bank IV 2026.

My role

  • Co-founder and FYP developer, responsible for the ML/AI architecture.
  • Built Roman Urdu / English text classification, from a TF-IDF + linear SVM baseline to XLM-RoBERTa, with a Gradio demo.
  • Built the local prototype's API and data handling, with tests for account isolation and validation.
  • Wrote the evaluation and data-audit tooling used by the research track.

Architecture

Two tracks that meet only when a model has been evaluated well enough to connect. That hasn't happened yet, and the site doesn't pretend otherwise.

Conceptual flow, simplified from the codebase.

  1. Research · text

    Roman Urdu / English classifiers: TF-IDF + linear SVM, then XLM-RoBERTa

  2. Research · face

    Facial-expression experiments on a public benchmark dataset

  3. Research · fusion

    A typed fusion step covering all seven modality combinations, with abstention

  4. Research · evaluation

    Macro-F1, per-class metrics, confusion matrix, Brier score, log loss, ECE

  5. Product · check-in

    Anonymous check-in in English or Roman Urdu. Nothing saved by default

  6. Product · data

    Saving needs an account, history permission and a per-check-in choice. Export and delete anytime

Key decisions

  • Nothing is saved unless the student asks for it, three separate times: an account, a history permission, and a choice on that check-in.
  • Research and training permissions are switched off in the prototype, so check-ins can't feed a dataset.
  • The fusion step is allowed to abstain. With weak or conflicting signals, saying nothing is better than a confident guess.

Validation

  • API and frontend tests cover account isolation and input validation.
  • A data-audit tool checks dataset rights, withdrawals and train/test overlap, and fails loudly when it finds a problem.

Limitations and next steps

  • No text, speech or face model is connected to the prototype yet.
  • The prototype runs locally on test content. It is not a public mental-health service and makes no clinical claims.
  • Next: connect an evaluated text model behind the abstention rule, then test it with real users under proper consent.