Saba Firdaus Ansaria
Machine Learning Engineer
I build ML systems that make it to production—not just the demo
From a commercialised fault-detection system (30% accuracy gain, live on a factory line) to containerized RAG chatbots and speech-AI research infrastructure — I take models from prototype through deployment, monitoring, and real users.
A short introduction
Machine Learning Engineer with an MSc in Robotics (First Class Distinction) and an MSc in Applied Mathematics, specialising in machine learning, LLMs/RAG, and computer vision.
Currently focused on speech AI — automatic speech recognition and speaker diarization — alongside broader ML engineering work.
Six years of experience communicating complex technical concepts to varied audiences — from research labs and engineering teams to classrooms and clients.
Right to work in the UK
Tools of the trade
Languages
ML & Deep Learning
Generative AI & LLMs
MLOps & Deployment
Tools & Practices
Mathematics & Statistics
Professional
Selected work
ASR & Speaker Diarization Research
Researching automatic speech recognition and diarization ranking systems at the University of Sheffield's School of Computer Science. Designing and implementing ranking metrics and systems, running experiments with Python and PyTorch on cloud/VM infrastructure.
Full-stack ML research discovery platform built from scratch. Fetches papers from arXiv across 7 research categories, scores them using a composite trend algorithm (citation velocity 40%, recency 30%, citation count 20%, keyword boost 10%), and clusters them into topic groups using TF-IDF and KMeans. Features an interactive topic heatmap, bookmark system, BibTeX citation copy, weekly email digest via Brevo, and nightly auto-fetch via APScheduler.
A menstrual health RAG chatbot rebuilt from a legacy LLaMA-2 prototype into a hybrid-retrieval system, deployed as three containerized microservices on Fly.io.
MSc thesis: multichannel CNN fusion model on NHS clinical data. Designed, trained and optimised for low-power wearables — from data pipeline to deployable inference.
Industrial Vision System
Commercialised real-time fault-detection system deployed on edge hardware. End-to-end delivery: model design, on-device optimisation, and production integration for a client factory line.
Improved accuracy 69.68% → 74.64% across CNN variants for spoken-digit recognition. Engineered spectrograms, filterbanks and MFCCs; built baseline, regularised and residual CNN architectures in TensorFlow.
Client-focused simulation with KPMG's Data, Analytics & Modelling team. Assessed data quality, performed EDA to identify high-value customer segments, and built interactive Tableau dashboards with actionable recommendations.
Career timeline
Research Intern
— The University of Sheffield- Built and maintained a self-hosted Codabench ASR Ranking Challenge platform end-to-end, configuring SMTP for service notifications and integrating MinIO object storage
- Developed a Kendall Tau scoring pipeline to rank and monitor submitted models, maintained as production infrastructure
- Produced formal technical documentation of the system's architecture for the supervising professor
CodabenchMinIOSMTPPythonASRFreelance ML Engineer
— Independent- Built and deployed a document-based RAG question-answering system using quantised Llama 2, LangChain, FAISS, Sentence Transformers and Streamlit, achieving sub-5-second CPU response times for on-device search across medical PDFs
- Scoped, built and shipped the project end-to-end with full autonomy; released open-source with reproducible documentation
RAGLangChainFAISSStreamlitLLaMA 2Software Consultant
— Spagiria Consulting- Developed a now-commercialised fault-detection system for a chemical production pipeline using multisensory data and image processing with Python, OpenCV and Raspberry Pi
- Reduced manual inspection time by 20% and increased fault-detection accuracy by 30%, taking the system from R&D through to live commercial deployment
- Worked directly with client stakeholders to scope requirements, validate data quality, and translate model output into actionable reporting for a non-technical audience
PythonOpenCVRaspberry PiEdge AIMathematics Teacher
- Designed and delivered an ICSE-aligned mathematics curriculum for 1,000+ students over six years, developing strong skills in simplifying complex quantitative material for varied audiences
Curriculum DesignCommunication
Academic background
M.Sc. Robotics
University of Sheffield
First Class Distinction
M.Sc. Applied Mathematics
B.Sc. Mathematics
Continued learning
- McKinsey Forward Programme (2025)
- Generative AI with LLMs — Coursera / DeepLearning.AI
- Machine Learning Specialisation — Coursera / Andrew Ng
Let's talk
Open to ML engineering roles, research collaborations, and consulting. The fastest way to reach me is by email — or via LinkedIn.
sabafirdaus.cs@gmail.com