Surbhi Lahoria
Computer Science graduate from The University of York with practical experience in iOS development and machine learning. Currently working as an iOS App Developer, designing mobile applications using Swift and SwiftUI for field operations teams. Expertise spans full-stack development, API integration, secure authentication, and CI/CD pipelines. Proven ability to collaborate with cross-functional teams, translating business requirements into technical solutions that improve operational efficiency.
Email Download CV (PDF) GitHub LinkedIn

RESEARCH

Residual-Bottleneck Autoencoder for Speech Anonymisation

Final Year Dissertation · BEng Computer Science · University of York

Supervised by Dr. Yuchen Zhao

May 2025 Dissertation

Abstract

Developed a novel privacy-preserving speech processing system using a residual-bottleneck autoencoder architecture to anonymise speaker identity while maintaining automatic speech recognition (ASR) accuracy. The system addresses critical privacy concerns in voice-assistant platforms by creating a lightweight, waveform-level anonymiser that sits seamlessly between raw audio input and existing ASR engines.

Technical Approach

  • Architecture: Frozen Wav2Vec2-base encoder with 256-dimensional residual bottleneck and U-Net-style decoder
  • Dataset: LibriSpeech train-clean-100 corpus (10 hours) for training, test-clean subset (128 utterances) for evaluation
  • Loss Function: Dual L₁ reconstruction loss combining time-domain waveform and Mel-spectrogram fidelity
  • Evaluation Metrics: Word Error Rate (WER) for utility, cosine similarity reduction for privacy using ECAPA-TDNN embeddings

Key Achievements

85.4%
Speaker Similarity Reduction
Exceeded 80% privacy threshold
14.9%
Final Word Error Rate
Below 20% usability threshold
0.041×
Real-time Factor
NVIDIA T4 GPU performance

Research Contributions

  • Demonstrated that strong speaker de-identification (85%+ reduction) and practical ASR usability can coexist in real-time
  • Validated the effectiveness of residual-bottleneck architecture for selective information suppression in speech processing
  • Provided comprehensive evaluation framework measuring both privacy and utility metrics on standard benchmarks
  • Achieved computational efficiency suitable for edge deployment without specialized hardware
Machine Learning:
PyTorch Wav2Vec2 ECAPA-TDNN Autoencoder
Signal Processing:
STFT Mel-Spectrograms Waveform Processing
Evaluation:
LibriSpeech Word Error Rate Speaker Verification
Full Dissertation (PDF) Code Repository

HIGHLIGHTS

iOS Field Operations App

Designed and built mobile application using Swift and SwiftUI for field operations teams. Implemented CSV-driven search functionality and workflow automation, reducing lookup time significantly. Features secure authentication, API integration, and deep-linking flows to improve operational efficiency.

SwiftUICombineFirebase AuthKeychain

Case study · Demo

Flowers-102 CNN Classification

Built and trained a compact CNN for 102-class flower recognition. Two conv blocks with ELU, BatchNorm, and MaxPool, plus dropout and FC layers; optimized with Adam and a cyclic learning-rate schedule. Heavy augmentation (flip/rotate/jitter/affine/grayscale) to improve generalization. Final test accuracy: 48.74%.

Python PyTorch ELU BatchNorm CyclicLR

Code · Report (PDF)

EXPERIENCE

iOS App Development Intern

GO4B UK

Jun 2025 – Present Remote

Designed and developed a multi-module SwiftUI mobile application to assist field engineers and site surveyors in identifying, configuring, and quoting industrial parts including buckets, bolts, chain conveyors, and electronics.

  • CSV-powered search interface: Built real-time part code lookup and auto-fill functionality using Excel datasets for enhanced operational efficiency
  • Dynamic forms with image previews: Implemented dimension inputs and visual components for various engineering parts and configurations
  • Modular SwiftUI architecture: Developed scalable app structure using @StateObject and @EnvironmentObject-driven state management
  • User authentication system: Designed secure onboarding and logout flows with planned Firebase Auth and Keychain integration
  • Step-based navigation flow: Created guided user journey through Product Identification → Site Survey → Engineering Inputs → Quotation Request
  • Cross-functional collaboration: Worked closely with engineering and marketing teams to refine UI/UX and ensure app meets field requirements for live deployment
SwiftUI Xcode Swift Data CSV Parsing Firebase Git Keychain

TECHNICAL SKILLS

💻

Programming Languages

Python Swift Java JavaScript
🧠

Machine Learning

PyTorch scikit-learn pandas NumPy
📱

Mobile Development

SwiftUI Combine iOS Development Xcode
🔧

Backend & DevOps

REST APIs Firebase CI/CD Git
📊

Data & Analytics

Statistical Analysis Power BI SQL Data Visualization
⚡

Development Practices

Code Quality Testing Agile Documentation

EDUCATION

BEng Computer Science

University of York

2022–2025 York, UK

Specialized in Machine Learning, Mobile Development, and Systems. Final project focused on Privacy-Preserving Speech Processing using real-time neural networks.

Key Areas: Machine Learning, iOS Development, Data Science, Software Engineering
Final Project: Privacy-Preserving Speech Processing with Neural Networks

CONTACT

Reach me at surbhilahoria@outlook.com. Seeking opportunities to apply technical skills and collaborative approach in challenging software development environments, with particular interest in mobile development, machine learning applications, and innovative technology solutions.