ZahidImran

An agent that queries, filters and ranks hotels on its own — then feeds a live BI dashboard.

• ▸Self-directed agentic workflow using the Google Maps API and function calling to rank hotels against complex constraints (price, rating, proximity).
• ▸ETL pipeline transforming raw API JSON into structured business intelligence for a real-time PowerBI dashboard.

Audio restoration at 14 dB SI-SNR and 2.25 PESQ, served in real time.

• ▸Fine-tuned a Transformer architecture for audio restoration, reaching 14 dB SI-SNR and 2.25 PESQ.
• ▸Deployed on Streamlit Cloud with an inference pipeline tuned for real-time audio streams at minimal latency.

A TTS voice that finally sounds right in Asian English accents.

• ▸Fine-tuned XTTS-v2 on a custom Asian-English accent dataset via transfer learning, improving prosody and naturalness for underrepresented dialects.
• ▸Built an automated audio preprocessing pipeline (clean, normalize, segment) for high-fidelity convergence.

Automating the dubbing of Urdu drama into Arabic, so the work reaches an audience it never had.

• ▸Built with the UrduX Lab team during the Grayhat internship: LLM-driven translation feeding an automated dubbing pipeline.
• ▸Paired with a Whisper model fine-tuned on Urdu that cut word error rate by 4%.

Reading emotion from speech at 72% accuracy across four model architectures.

• ▸Compared CNN, ANN, LSTM and hybrid CNN+LSTM models on the same speech-emotion task.
• ▸The hybrid architecture reached 72% accuracy — the strongest of the four.

A full training-to-deployment loop that runs itself on every commit.

• ▸End-to-end MLOps pipeline with GitHub Actions and Docker automating collection, preprocessing, training and deployment.
• ▸MLflow for experiment tracking and model registry, DVC for data versioning — every run reproducible.