Emily Perry

Senior Machine Learning Researcher, Fourthline

Technical highlights

• Designed multimodal fraud detection features using selfie videos and identity document imagery
• Generated semantic embeddings from visual identity verification data
• Explored vision-language prompting approaches for fraud-related feature extraction
• Evaluated dimensionality reduction, clustering, and MLPs
• Developed feature selection strategies for integration with mature tree-based models
• Combined multimodal representation learning with production fraud detection systems
• Improved fraud detection recall by 8% in a regulated KYC environment

Fourthline's fraud detection platform relied on a mature tree-based machine learning system that had been refined over many years and already incorporated a large number of engineered features. Rather than rebuilding existing work, I focused on identifying new information sources that could provide additional predictive signal beyond the current feature set.

I proposed augmenting the fraud detection pipeline with multimodal representations derived directly from customer identity verification data, including selfie videos and identity document imagery. The objective was to capture subtle behavioural and visual patterns that were not represented in the existing structured features.

To achieve this, I developed a feature generation framework based on modern embedding models. Identity documents and selfie video frames were transformed into dense vector representations, creating a rich semantic description of each verification attempt. A major challenge was determining how to convert these high-dimensional embeddings into features that could effectively complement an existing tree-based production system.

I explored a wide range of approaches, including prompt-based embedding extraction, dimensionality reduction techniques, clustering methods, neural feature transformations, and learned latent representations. Experiments included using vision-language models to evaluate concepts such as suspicious or potentially fraudulent documents, alongside unsupervised methods designed to identify anomalous patterns within the embedding space.

Rather than relying on a single modelling approach, I systematically evaluated multiple feature-generation strategies and selected the most informative signals for integration into the production fraud detection model. This hybrid approach combined the strengths of modern multimodal representation learning with the robustness and interpretability of the existing tree-based architecture.

The resulting system delivered an 8% improvement in fraud detection recall, significantly increasing the platform's ability to identify fraudulent applications while preserving the strengths of the existing production pipeline.

Senior Machine Learning Researcher, Fourthline

Technical highlights

• Redesigned an end-to-end KYC automation decision framework
• Developed risk-scoring methodologies combining signals from multiple verification stages
• Created client-specific automation strategies aligned to differing risk appetites
• Built feedback-driven threshold optimisation mechanisms for dynamic risk management
• Analysed fraud-performance trade-offs across large-scale production systems
• Increased automated verification rates by 4% with minimal impact on fraud exposure
• Delivered measurable operational efficiency improvements in regulated financial environments

Fourthline's identity verification platform processes large volumes of Know Your Customer (KYC) cases, balancing two competing objectives: maximising automation while maintaining stringent fraud controls.

The original scope of this project was relatively narrow: identify customers and regions where automation had not been enabled and determine whether additional automation opportunities existed. After reviewing the end-to-end verification process, I identified a significantly larger opportunity. Rather than focusing on individual automation gaps, I redesigned the decision framework underpinning the entire automation pipeline.

I conducted a detailed analysis of verification outcomes across the full sequence of identity checks, examining where cases were exiting the automation process and whether those decisions were consistent with the underlying risk profile of each application. This revealed opportunities to improve automation rates without materially increasing fraud exposure.

To address this, I developed a risk-based decision framework that evaluated the cumulative evidence produced throughout the verification journey and estimated the overall risk associated with each case. The framework incorporated client-specific risk appetites, allowing different organisations to make automation decisions that reflected their own regulatory and operational requirements.

A key challenge was balancing automation gains against fraud risk - and the appetite of the clients. Rather than relying on static thresholds, I designed a feedback-driven optimisation system that enabled decision boundaries to be adjusted dynamically as fraud patterns and business requirements evolved. This provided significantly greater operational flexibility while maintaining strong control over risk exposure.

The redesigned framework increased automated verification rates by approximately 4%, delivering substantial operational savings while producing only a negligible increase in fraudulent cases entering the automated flow (approximately 0.01%). The work demonstrated how machine learning and decision science can be applied not simply to improve model accuracy, but to optimise an entire production decision system.

Lead Machine Learning Researcher, EDITED

Technical highlights

• Built large-scale time series forecasting systems across billions of retail observations
• Developed statistical normalisation techniques to remove catalogue growth and seasonal effects
• Modelled expected behaviour using historical trend forecasting
• Identified emerging trends through deviation analysis and anomaly detection
• Generated early indicators for colours, materials, patterns, silhouettes, and garment categories
• Delivered production analytics supporting strategic merchandising and buying decisions

One of EDITED's core challenges was identifying emerging fashion trends before they became mainstream. With billions of products collected from thousands of global retailers, the objective was not simply to forecast demand, but to distinguish genuine changes in consumer behaviour from normal seasonal fluctuations and catalogue growth.

I developed large-scale time series analytics that transformed raw retail inventory data into leading indicators of emerging fashion trends across colours, patterns, materials, garment types, and product attributes.

A significant challenge was separating true market signals from the natural dynamics of the fashion industry. The volume of products entering the platform continually increased over time, while strong seasonal effects meant many trends repeated predictably each year. To overcome this, I designed statistical normalisation pipelines that accounted for long-term catalogue growth, historical seasonality, and recurring annual patterns before any trend analysis was performed.

Using historical behaviour as a baseline, I developed forecasting models to estimate the expected trajectory of each fashion attribute. Products or attributes that significantly exceeded their expected behaviour—such as colours, patterns, or garment styles growing substantially faster than historical trends predicted—were identified as emerging trends. Conversely, underperforming categories were detected early, allowing customers to identify declining demand before it became apparent in traditional retail reporting.

The resulting platform enabled customers to move beyond descriptive analytics and identify fashion movements while they were still developing. These insights became a key component of EDITED's fashion intelligence offering, supporting buying, merchandising, and product strategy decisions for global retailers.

Lead Machine Learning Researcher, EDITED

Leadership highlights

• Established technical standards for experimentation, deployment, and model evaluation
• Acted as technical lead across recommendation, classification, forecasting, and computer vision programmes
• Partnered with senior stakeholders to align machine learning strategy with business objectives
• Balanced hands-on technical leadership with team management responsibilities
• Managed teams career progression

Alongside leading the technical direction of EDITED's machine learning initiatives, I managed a team of data scientists and machine learning engineers responsible for delivering production systems across classification, recommendation, forecasting, and computer vision.

In a time when the team lacked formal management, I took responsiblity for career progression, team performance, and objective setting, while implementing teamwide mentoring and technical coaching. I worked closely with team members to support both their technical growth and long-term career progression while maintaining a high-performing and collaborative engineering culture.

At the same time, I continued to serve as the technical lead for the machine learning function. This involved defining architectural direction, reviewing model designs, establishing best practices for experimentation and deployment, and guiding major technical decisions across multiple concurrent projects.

A key focus of the role was balancing delivery with technical excellence. I partnered with product, commercial, and engineering stakeholders to prioritise machine learning investments while ensuring systems remained scalable, maintainable, and aligned with long-term business objectives.

This combination of technical leadership and people management allowed me to scale my impact beyond individual contributions, enabling the successful delivery of multiple machine learning platforms while supporting the growth of the wider team.

Lead Machine Learning Researcher, EDITED

Technical highlights

• Built multimodal recommendation and retrieval systems across a catalogue of more than one billion products
• Developed shared text and image embedding architectures for semantic similarity search
• Designed embedding compression strategies to reduce storage and retrieval costs
• Implemented large-scale approximate nearest-neighbour search using HNSW graph algorithms
• Developed multi-stage retrieval and re-ranking pipelines for improved recommendation precision
• Built image-based search capabilities allowing users to find similar products from uploaded photographs
• Integrated LLM-based summarisation to explain product similarities and improve discovery workflows
• Delivered production systems supporting real-time recommendation and visual search

One of the most challenging problems at EDITED was helping users discover visually and stylistically similar products within a catalogue containing more than one billion retail products collected from thousands of global retailers.

Traditional recommendation approaches were not applicable. There was no collaborative filtering signal, no customer interaction history, and no labelled training dataset defining product similarity. The problem therefore became one of large-scale representation learning and retrieval: how can we identify products that are genuinely similar when all we have is their content?

I designed and led the development of a multimodal recommendation platform that transformed both product imagery and textual descriptions into shared semantic embeddings. These representations captured style, silhouette, colour, material, and other product characteristics, enabling similarity search across an extremely large and constantly evolving product universe.

A major engineering challenge was making retrieval computationally feasible at scale. To support fast nearest-neighbour search across billions of products, I developed embedding compression pipelines and implemented approximate nearest-neighbour retrieval using Hierarchical Navigable Small World (HNSW) graphs. This reduced search latency dramatically while maintaining recommendation quality.

The retrieval architecture was designed as a multi-stage system. The HNSW index generated a candidate set of potentially relevant products, which were then passed through additional filtering, ranking, and machine learning models designed to identify the most accurate stylistic matches. This approach balanced retrieval speed with recommendation precision and enabled production deployment at massive scale.

The platform evolved beyond internal product discovery into a customer-facing visual search capability. Users could upload their own product images, which were converted into embeddings and matched against the global catalogue in real time. Large language models were subsequently integrated to generate natural-language summaries of visual characteristics and explain recommendation results, creating a more intuitive search and discovery experience.

The resulting system became a core product capability, enabling large-scale similarity search, competitive product analysis, trend discovery, and customer-facing visual recommendation experiences across the platform.

Senior Machine Learning Researcher, EDITED

Technical highlights

• Built large-scale multimodal classification systems using text and image data
• Developed hierarchical classification models across thousands of product categories and attributes
• Designed shared embedding architectures to support multiple downstream prediction tasks
• Combined deep learning representations with gradient-boosted models for efficient production inference
• Managed large-scale machine learning infrastructure and deployment pipelines
• Tackled severe class imbalance and long-tail classification challenges
• Partnered with annotation teams to improve training data quality and taxonomy consistency
• Enabled automated attribution across millions of retail products

EDITED's platform relies on accurately understanding and structuring fashion, beauty, and homeware products from thousands of global retailers. I worked on the development of large-scale machine learning systems designed to automatically classify products across a complex hierarchical taxonomy, ranging from broad categories such as apparel, beauty, and homeware through to highly specific attributes including style, fit, sleeve length, material, colour, and sizing.

The challenge extended far beyond a traditional classification problem. The taxonomy contained thousands of highly imbalanced classes, many with limited training examples, while product descriptions and imagery varied significantly between retailers. Building a scalable and maintainable system required solving challenges in data quality, representation learning, infrastructure design, and model architecture.

Initially, I developed specialised classification systems that independently leveraged textual and visual product information. These models were used throughout the hierarchy to classify products at multiple levels of granularity, enabling the automated enrichment of millions of products across the platform.

As the platform matured, I led the transition towards a more generalised multimodal architecture. Rather than maintaining separate models for each classification task, I developed shared embedding representations from text and image data that captured product semantics across the taxonomy. These embeddings became the foundation for a wide range of downstream tasks, allowing lightweight models such as XGBoost to efficiently solve individual classification problems while benefiting from a common learned representation.

Alongside the modelling work, I was responsible for managing the large-scale technical infrastructure supporting the platform. This included maintaining and evolving a complex monolithic machine learning codebase, managing dependencies across multiple modelling pipelines, and ensuring reliable deployment and monitoring across a growing suite of production systems.

A major focus of the project was addressing extreme class imbalance and maintaining model performance across both common and rare product categories. Working closely with EDITED's in-house annotation teams, I developed data quality processes, evaluation frameworks, and modelling approaches designed to balance bias and variance while ensuring robust performance throughout the hierarchy.

The resulting platform became a core component of EDITED's product intelligence engine, enabling scalable product attribution and powering downstream analytics, forecasting, recommendation, and trend detection systems.

Senior Data Scientist, Deliveroo

Technical highlights

• Developed real-time rider travel-time prediction models for logistics optimisation
• Integrated rider, route, weather, temporal, and operational signals into predictive systems
• Conducted large-scale A/B testing to evaluate production performance
• Analysed interactions between machine learning predictions and optimisation algorithms
• Identified and mitigated feedback loops affecting rider allocation
• Evaluated system-level impacts beyond traditional predictive accuracy metrics

A critical component of Deliveroo's logistics platform is accurately estimating how long a rider will take to travel between locations. These predictions influence customer delivery estimates, rider assignment decisions, and overall network efficiency.

I developed machine learning models to predict rider travel times in real time, combining behavioural, environmental, and operational signals to estimate journey duration across a highly dynamic urban delivery network.

The models incorporated a wide range of features including transport mode, historical rider performance, route characteristics, weather conditions, time-of-day effects, and local demand patterns. The objective was to generate accurate travel-time estimates that could be used by Deliveroo's dispatch and optimisation systems to make better operational decisions.

The project achieved strong predictive performance and was extensively evaluated through large-scale online experimentation. However, deployment revealed an important challenge that was not apparent in offline model evaluation.

Because predictions were generated at the individual rider level and fed into downstream optimisation systems, the model began influencing how work was allocated across the rider network. High-performing riders were increasingly favoured by the dispatch algorithm, while other riders received fewer opportunities to complete deliveries. This created a feedback loop where the optimisation system reinforced historical performance differences and unintentionally encouraged undesirable marketplace dynamics.

Identifying and addressing these effects required looking beyond traditional machine learning metrics and understanding the interaction between predictive models and the broader decision-making system. The experience fundamentally shaped my approach to production machine learning, highlighting the importance of fairness, incentive structures, feedback loops, and responsible deployment when models directly influence human behaviour.

The project ultimately became an important lesson in the difference between building an accurate model and building an effective, equitable decision system.

Data Scientist, Deliveroo

Technical highlights

• Designed and delivered Deliveroo's first personalised recommendation system
• Developed ranking algorithms using historical customer behaviour and collaborative signals
• Combined user preferences, item popularity, and business objectives into a unified scoring framework
• Built recommendation pipelines suitable for production deployment at scale
• Worked across product, engineering, and data teams to integrate recommendations into the customer journey
• Established foundational recommendation infrastructure for future personalisation initiatives

As Deliveroo rapidly expanded, the platform lacked a personalised recommendation capability to help customers discover relevant menu items and restaurants. I led the development of the company's first recommendation systems, designing and delivering a production-ready solution from concept to deployment.

The challenge was to create a ranking system capable of personalising menu recommendations despite limited existing infrastructure and evolving product requirements. Working in a fast-paced environment, I developed a recommendation framework that combined customer behaviour, collaborative signals, and business objectives to improve the relevance of menu rankings.

The system incorporated multiple sources of information, including customers' historical ordering patterns, preferences inferred from similar users, item popularity, and operational business metrics. By combining these signals into a unified ranking framework, the model was able to surface menu items that were both personally relevant and commercially effective.

A key aspect of the project was balancing recommendation quality with practical deployment constraints. With no existing recommendation platform in place, I designed the initial architecture, developed the modelling approach, and worked closely with engineering and product teams to integrate recommendations directly into the customer experience.

The result was a scalable recommendation capability that established the foundations for personalised discovery on the platform and demonstrated the value of machine learning-driven ranking systems within the business.

Data Scientist, Deliveroo

Technical highlights

• Developed real-time machine learning models for restaurant preparation time prediction
• Modelled complex interactions between restaurant behaviour, order composition, and temporal factors
• Engineered features from historical operational and customer order data
• Built forecasting systems serving production logistics workflows
• Designed and analysed large-scale A/B experiments to measure business impact
• Improved ETA accuracy and operational decision-making across the delivery network

One of the key challenges in food delivery logistics is accurately predicting how long a restaurant will take to prepare an order. Small forecasting errors can compound throughout the delivery process, leading to missed customer expectations, inefficient rider allocation, and increased delivery times.

I led the development of machine learning models to predict restaurant food preparation times in real time, providing a critical input into Deliveroo's delivery estimation and logistics systems.

The challenge was highly dynamic. Preparation times varied significantly between restaurants and depended on a combination of factors including menu items ordered, order complexity, restaurant workload, time of day, day of week, and historical operating performance. I developed predictive models that incorporated these behavioural and operational signals to generate accurate preparation time estimates at the point an order was placed.

A significant part of the project involved feature engineering and understanding the operational drivers of restaurant performance at scale. By modelling the interactions between order characteristics and restaurant-specific behaviour, the system was able to adapt predictions to the unique operating patterns of thousands of restaurants across the platform.

To validate business impact, I designed and analysed large-scale A/B tests, measuring the effect of improved preparation time estimates on delivery accuracy, operational efficiency, and customer experience. The resulting models became a core component of Deliveroo's logistics optimisation infrastructure, helping improve delivery predictions and support more efficient rider assignment decisions.

Data Scientist, Telefonica Smart Steps (Secondment)

Technical highlights

• Analysed datasets containing billions of geospatial records
• Developed scalable SQL and PostgreSQL analytical pipelines
• Built origin-destination and mobility models from anonymised location traces
• Designed methodologies for privacy-preserving population analysis
• Managed end-to-end consultancy engagements independently
• Presented findings and recommendations directly to senior stakeholders

During a secondment with Telefonica's Smart Steps division, I led a portfolio of mobility analytics projects built on anonymised location data from millions of mobile devices across the UK. The underlying dataset contained billions of latitude-longitude observations, providing a unique opportunity to understand population movement patterns at national scale while maintaining strict privacy and anonymity requirements.

Working independently as both technical lead and consultant, I owned the full project lifecycle from stakeholder discovery through to delivery and executive presentation. Each engagement required translating often loosely defined business questions into robust analytical frameworks and actionable insights.

Key challenges included transforming raw geospatial telemetry into meaningful behavioural signals. I designed and implemented large-scale analytical pipelines using complex PostgreSQL querying and aggregation techniques to infer:

• Home and work location distributions
• Commuting and travel behaviour
• Seasonal and holiday migration patterns
• Visitor origin-destination flows
• Regional mobility trends and catchment analysis
• Population movement models for events and infrastructure planning

The work supported organisations seeking to better understand customer movement, tourism patterns, transport demand, and regional economic activity. By combining large-scale data engineering with statistical analysis and geospatial modelling, I was able to deliver evidence-based recommendations that informed strategic planning decisions across a range of sectors.

Data Scientist, Arriva Rail North (Secondment)

Technical highlights

• Developed statistical models to forecast franchise operating costs
• Applied queueing theory to model passenger congestion and service capacity
• Built predictive demand and crowd-flow analytics for rail networks
• Quantified operational risk across multiple service scenarios
• Combined forecasting, optimisation, and financial modelling techniques
• Delivered analytical outputs to support high-value commercial decisions

As part of a secondment to Arriva Rail, I worked on the bid team responsible for securing the Northern Rail franchise, developing quantitative models to forecast the operational and financial performance of the proposed rail network.

The challenge was to accurately estimate the long-term costs, capacity requirements, and operational risks associated with a franchise serving millions of passenger journeys annually. Decisions made during the bidding process had significant commercial implications, requiring forecasts that were both statistically robust and operationally realistic.

I developed and applied advanced statistical and operational research techniques to model passenger demand, network congestion, and service delivery costs. This included combining predictive analytics with queueing theory and capacity modelling to understand how passenger flows would interact with station infrastructure, train services, and operational constraints.

The work involved analysing large transportation datasets to forecast demand patterns, identify potential bottlenecks, and quantify the financial impact of different service scenarios. These models provided evidence-based estimates of operating costs and supported strategic decision-making throughout the bid process.

By integrating statistical forecasting, optimisation techniques, and transport operations research, I helped deliver a comprehensive analytical framework used to evaluate the viability and competitiveness of the franchise proposal.