Neo4J Recommender System

Social Media

---

context

I was working for a pre-launch social media and concierge app as the sole data person. My responsibilities included designing the data model, managing the analytics infrastructure in AWS using Terraform, and setting up pipelines to ingest, clean, and store data.

task

The main page of the app featured a feed of experiences (e.g., hotels, restaurants, events) that users could explore and book. It was crucial that the content matched each user’s browsing habits, necessitating a personalised feed. We aimed for a low-touch onboarding process without requiring users to specify their preferences.

approach

I chose to use a graph network for the recommender system, leveraging its strength in representing relationships and interactions between entities. In our case, the nodes were users and experiences, and the edges represented interactions (view, book, save, review). I manually set weights for these interactions based on perceived user intent, viewing an experience counted as 1, saving as 5, and booking as 10.

outcome

Although the app has not launched yet, the infrastructure is in place and documented in Terraform, with scripts stored in GitHub. I provided extensive documentation and walkthroughs for the development team to facilitate post-launch activation.

• Client: Pre-launch Social Media Startup
• Tech: Neo4J, AWS (Neptune, Kinesis, Lambda)

Close

Building a Data Warehouse and Setting up Data Pipelines in AWS

Tech Platform

---

context

I joined a scale-up tech platform where major brands crowd-source video content for social media. Despite nearly a decade of operation, the company lacked a unified data view or a single source of truth.

task

I was tasked with extracting data from various sources and loading it into a Redshift database. After landing the data, I needed to clean and transform it to support the Marketing, Finance, and Product teams.

approach

Reporting directly to the head of operations, I utilised AWS Glue for its user-friendly interface, which provided transparency for non-coders. Despite its slower setup compared to pure Python scripts, it facilitated better understanding and oversight of the ETL processes.

outcome

I successfully ingested, modelled, and built reports using data from multiple sources, including Mixpanel, backend databases, Airtable, Google Sheets, Jira, GCP APIs, TikTok, Meta, and HuggingFace. These reports enabled various teams to make informed decisions, enhancing overall operational efficiency.

• Client: Content Platform Scale-up
• Tech: AWS Glue

Close

Migrating a Process from Jenkins to Snowflake for a Financial Services Company

Financial Services

---

context

I joined a 200-year-old asset management company to provide maternity cover and support for their data and insights team.

task

My primary task was to migrate a complex process for preparing a dataset from Jenkins to Snowflake. The dataset detailed geographical revenue of companies, aiding portfolio managers in adjusting and rebalancing holdings based on macro trends.

approach

I mapped out the 14 existing Jenkins jobs, identifying key steps to simplify the process. This involved collecting up-to-date portfolio data, revenue exposure data, mapping various identifiers (ISIN, SEDOL, Ticker Code) to the GeoRev dataset, and aggregating the data at the portfolio level. I leveraged Snowflake’s materialised views for daily data preparation, eliminating reliance on Jenkins.

outcome

I completed the migration, testing, and documentation in four weeks—significantly faster than the three-month estimate given by the internal data engineering team. The new process also included a historical data view, enabling asset managers to track portfolio exposure changes over time.

• Client: Pre-launch Social Media Startup
• Tech: Jenkins, Snowflake

Close

Product Analysis for a Food Order and Delivery Platform

Tech Platform

---

context

I joined one of the UK’s largest tech companies, enabling users to find and order food from local restaurants. The company had expanded rapidly through multiple acquisitions.

task

My task was twofold:to determine the overlap between users of the acquired companies and the main platform, and to analyse where users were dropping off during the migration process from the acquired platform (Hungry House) to the main platform (Just Eat).

approach

While a team of BI analysts handled the first question, I focused on the user migration funnel. I mapped the ideal user journey and identified key data points in Google Analytics via BigQuery. Notably, I discovered that the banner prompting users to switch platforms was more effective on mobile than desktop, due to its prominence on the screen.

outcome

Based on my recommendation, the banner size on desktop was increased, resulting in a significant rise in click-through rates and approximately ten thousand additional orders. This highlighted the importance of reevaluating seemingly obvious issues from a fresh perspective.

• Client: Food Delivery Platform
• Tech: BigQuery, Google Analytics, Tableau

Close