Tabinda Sandhu

Computer Science Graduate specializing in Urdu Speech AI and Business Automation

Based in: Lahore, PakistanMost recently: Business Automation Intern, Systems Ltd.

Hire me

A FAST-NUCES graduate with extensive experience in fine-tuning Whisper models and building voice agents for Urdu speech recognition. She has a proven track record in automating business workflows and developing full-stack applications using Python, FastAPI, and React.

Experience

Business Automation Intern

Systems Ltd.LahoreinternshipNov 2025 – Mar 2026

Summary

Built a customer service agent in Copilot Studio and Power Automate that read customer emails, sorted each query and triggered a reply. Built and maintained UiPath and Power Automate workflows that connected business systems through APIs. Traced and fixed failing workflows, which made the automations more reliable in daily use.

What I did

  • Improved the dataset to ensure more accurate category classification.
  • Debugged the system by testing on more than 100 emails daily.

Results

  • Reduced classification failures to only 2-3 emails daily by fixing the dataset and prompts.
  • Tested the system on an estimated volume of more than 500 emails per day.
  • Collaborated with a team of 3 developers and a QA team for daily system testing.
Copilot StudioPower AutomateUiPathAPIs

Skills

Technical

Hugging Face Transformers
PyTorchPyTorch
Whisper
PythonPython
JavaScriptJavaScript
ReactReact
REST APIs
Gemini API
RAG
ChromaDB
FastAPIFastAPI
GitGit
TypeScriptTypeScript
Hugging Face Hub
DjangoDjango
SQLSQL
TensorFlow/KerasTensorFlow/Keras
Scikit-learn
ViteVite
PostgreSQLPostgreSQL
SQLiteSQLite
LinuxLinux
Copilot Studio
Power Automate
Logistic Regression

Languages

Urdu
English

Projects

Support Ticket Classifier MVP

personal
Built an MVP that sorts customer issues into categories and generates summaries with suggested replies.
Logistic RegressionGemini

Urdu Speech Recognition: Whisper Benchmark and Fine-tuning

personal
Benchmarked Whisper base, small and large-v3-turbo on 299 Urdu FLEURS test clips (WER 51.5%, 36.7%, 22.1%). Fine-tuned Whisper-small on about 7 hours of Urdu speech: WER 36.7% → 28.6%, CER 13.7% → 10.1%. Added Urdu text normalisation, forced Urdu output after drift into Hindi script, and a repetition-loop guard. Converted it to int8 with CTranslate2 to run on a laptop CPU, and published both versions on Hugging Face.
PythonHugging Face TransformersPyTorchCTranslate2
Visit the site

Urdu COD Order-Confirmation Voice Agent

personal
Browser-based voice agent that talks to cash-on-delivery customers in Urdu to confirm, cancel or update an order. Gemini maps each customer reply to one of 8 intents, but a state machine in code makes the final decision, so the model can never confirm or cancel an order by itself. Answers store policy questions with RAG over the company's documents (16 of 17 correct at Hit@1 on my retrieval tests) and only answers when the retrieved text supports it. Speaks through Uplift AI's Orator TTS with caching and two fallback voices, plus my Urdu number normaliser. Built a 17-scenario automated test suite and logged ASR, LLM and TTS latency for every turn.
FastAPIReactTypeScriptfaster-whisperGemini APIChromaDBSQLiteUplift AI Orator TTS

Urdu Tune: Speech Data Collection and Tuning Tool

personal
Web app where a non-expert records Urdu clips in the browser, gets a draft transcript from my fine-tuned Whisper, and corrects it using keyboard shortcuts. Uses the corrections to measure the model's WER on the user's own audio before any training (about 30% so far). Splits data into train and test by speaker, warns about sentence overlap and too little audio, and exports a versioned private dataset to Hugging Face.
ReactTypeScriptFastAPISQLiteHugging Face Hub

CTC Speech Recognition (English)

academic
Trained a CNN + 5-layer bidirectional GRU model (2.8M parameters) with CTC loss on LJSpeech (13,100 clips), reaching 24.23% validation WER with greedy decoding.
PythonTensorFlowKerasCNNBiGRU

Auditra: Compliance Monitoring System

academic
Full-stack app that flags unusual access in system logs using anomaly detection and rule-based checks.
PythonDjangoReactPostgreSQLScikit-learnIsolation ForestOne-Class SVMz-score

Education

Bachelor of Science

National University of Computer and Emerging Sciences (FAST-NUCES)Computer Science2022 – 2026

Recognition

Certifications

Build an Agent in a Day

Microsoft

Essentials of Prompt Engineering

AWS

Google IT Automation with Python

Google
Hire TabindaReach out about a role, a contract or a conversation.For recruiters

Tabinda's twin is AI, it can make mistakes.

More people on Kanvis

Browse more peopleCreate your own Kanvis