Bhavesh Nathwani — Senior Data Analyst

02 · Experience

Canopy

Founding Data Hire & Senior Data Analyst

Oct 2020 – Jan 2026 · London · 5.5 years

Oversaw all end-to-end data analytics, infrastructure and insights across a 5.5-year tenure, supporting the business through multiple funding rounds, a strategic competitor acquisition and growth from 3,000 to over 1 million users.

Senior Data Analyst Mar 2024 – Jan 2026

Promoted to full ownership of data strategy, infrastructure, engineering and analytics. Served as core data stakeholder during the acquisition of Canopy's largest competitor.

Published the data architecture and methodologies behind Canopy's Rental Affordability Index, cited 400+ times by the BBC, Sky News and The Independent. Significantly drove brand authority and organic reach.
Led post-acquisition data integration: mapped and reconciled highly complex, disparate operational datasets from both companies to create a single source of truth for company-wide KPIs and reporting.
Overhauled core data infrastructure via a major migration from Google BigQuery and AWS Redshift to PostgreSQL, reducing pipeline processing times by 90%.
Drove adoption of the company's self-serve analytics framework, empowering cross-functional teams to build their own reports, cutting ad-hoc data requests by 75%+.
Mentored a Junior Data Analyst through to a Principal Data Analyst contract at the BBC. Established company-wide data governance, integrity checkpoints and automated reconciliation frameworks.

🏆 British Data Awards 2025 Finalist

Data Analyst Oct 2020 – Mar 2024

Appointed as the first data hire, architecting and launching Canopy's central data function from scratch.

Evaluated and overhauled the BI tech stack, migrating from vendor-implemented AWS QuickSight to Metabase, a more intuitive, cost-effective and scalable solution.
Built scalable reporting infrastructure, automating weekly, monthly and quarterly operational insight loops that removed manual bottlenecks and improved leadership forecasting accuracy.
Introduced predictive modelling and cohort analysis to evaluate user behaviour, identify churn risks and surface expansion opportunities for product and marketing teams.

SQL

Python

PostgreSQL

BigQuery

Metabase

Redshift

Google Cloud

AWS

Dataform

n8n

Power Automate

Apps Script

Google Sheets

Excel

Automation

Claude

03 · Experience

Webalytix

May 2019 – Sep 2020 · London · 1.5 years

Appointed to support clients in leveraging data for strategic decision-making, bridging the gap between raw data and actionable insights. Operated in a consultant capacity, working directly with stakeholders to deliver tailored analytics solutions.

Graduate Insight Analyst May 2019 – Sep 2020

Led end-to-end data projects from scoping and requirements gathering through to delivery, ensuring high-impact, client-ready outputs.
Developed and automated data workflows using SAS, SQL (SQL Server, MySQL, Snowflake) and Python, incorporating basic machine learning models.
Designed interactive dashboards and visualisations in Tableau, Excel and Python, transforming complex datasets into clear, actionable insights.
Delivered well-documented analytical outputs ensuring clients had a thorough understanding of methodology and outcomes.

SQL

Python

Tableau

SAS

Snowflake

04 · Education

Education

MSc Applied Data Science & AI

Oct 2024 – May 2026 · 1.5 years

OPIT – Open Institute of Technology

98%Coursework GPA

84%Dissertation GPA

Academic ExcellenceFinal Grade

Dissertation: My dissertation focused on building a hybrid human-AI system to automatically discover undeclared foreign key relationships in large-scale relational databases, a problem that sits at the heart of AI-powered data discovery and database automation.

Read about my dissertation →

Modules:

Applied Artificial Intelligence Applied Machine Learning Big Data and Cloud Computing Infrastructure Project Management Python Programming for Data Science Business Problem Solving Applications in Data Science and Artificial Intelligence Business Communication Ethics and Regulations of Artificial Intelligence Research Methods and Tools

BSc Mathematics with Honours

Sep 2015 – Sep 2018 · 3 years

The University of Manchester

Hindu Society · Mathematics Society · PASS Leader · UoM Buddy Scheme

Year 1:

Mathematical Workshop Sets, Numbers and Functions Calculus and Vectors Probability 1 Linear Algebra Calculus and Applications Sequences and Series Introduction to Statistics

Year 2:

Fundamentals of Financial Reporting Fundamentals of Management Accounting Real and Complex Analysis Calculus of Several Variables Algebraic Structures 1 Algebraic Structures 2 PDEs and Vector Calculus Discrete Mathematics Introduction to Financial Mathematics

Year 3:

Probability 2 Foundations of Modern Probability Statistical Methods Group Theory Hyperbolic Geometry Algebraic Geometry Mathematical Logic Green's Functions & Integral Equations Wave Motion Symmetry in Nature Matrix Analysis Martingales with Applications Markov Processes

05 · Technical Tech Stack

Tools & Technologies

SQL

Python

PostgreSQL

BigQuery

Metabase

Redshift

Google Cloud

AWS

Dataform

n8n

Power Automate

Apps Script

Google Sheets

Excel

Automation

SQLAlchemy

OpenAI API

SQLite

pandas

scikit-learn

Rich

Claude

MSc Dissertation · 2026

98%Coursework GPA

84%Dissertation GPA

Academic ExcellenceFinal Grade

Evaluating Hybrid Human-AI Systems for Discovering Undeclared Foreign Key Relationships in Large-Scale Relational Databases

MSc Applied Data Science & AI · OPIT – Open Institute of Technology ↗ · Supervised by Professor Lokesh Vij

92.64%

F1 Score

238M

Rows Processed

9 mins

Full Pipeline Runtime

368

Tables

310

Known Foreign Keys

94.02%

Precision

91.29%

Recall

+15

percentage point F1 gain vs. name matching alone

Built entirely in Python, evaluated on real open-source data.

Python

PostgreSQL

SQLAlchemy

OpenAI API (GPT-4)

SQLite

pandas

scikit-learn

Rich

1 · The Problem

Foreign keys are missing. Nobody notices until it's too late.

Example of a real-world database with missing foreign key documentation

Foreign key relationships are the backbone of any relational database. In enterprise environments, they are frequently dropped for performance reasons, lost during migrations, or simply never documented in legacy systems.

A database with 1,800 columns has over 3 million possible column pair combinations. Finding the meaningful ones manually is not realistic. My dissertation asked whether a system could do it automatically, at scale, with enough accuracy to be trusted.

2 · The Solution

A six-step pipeline combining statistics, naming patterns and human expertise.

1

Schema Reading

Connects to the database and reads structure: tables, columns, data types, indexes. Foreign key information is deliberately excluded, simulating the real-world scenario where it is absent.

2

Column Profiling

Every column is sampled to build a statistical profile: distinct value counts, null rates, and whether it looks like a key column.

3

Candidate Generation

Identifies column pairs that could be foreign key relationships using naming patterns and structural rules to narrow millions of combinations.

4

Multi-Signal Scoring

Each candidate is scored via two signals: heuristic (name similarity, type compatibility) and value containment (whether values in one column appear in the other). Combined into a single confidence score.

5

Human Review

Uncertain candidates are surfaced with full context: confidence score, signal breakdown, sample values side by side, and a plain-English GPT-4 explanation. The reviewer confirms, rejects, or marks as uncertain.

6

Knowledge Base

Every human decision is stored and used to improve future runs, adjusting signal weights and learning new naming patterns over time.

3 · Human Review Interface

Every decision is explainable. No black boxes.

Human Review Interface: CLI screenshot

The centrepiece of the system is a CLI tool that presents uncertain candidates to a reviewer with full context: confidence score, signal breakdown, sample values from both columns side by side, and a plain-English explanation generated by GPT-4.

Reviewers are never shown a black box score. Every decision is fully explainable, with GPT-4 generating a plain-English rationale for each candidate surfaced for review.

4 · Results

92.64% F1: over 15 percentage points above either signal alone.

Configuration	F1 Score
Name matching only	76.90%
Value analysis only	78.61%
Combined (final system)	92.64%

The system was evaluated against 310 known foreign key relationships in MusicBrainz, a real open-source music database with 368 tables and 238 million rows. Testing each component in isolation revealed how much each signal contributed.

Neither signal alone was sufficient. Combined, they produced a result over 15 percentage points higher than either individually.

5 · Key Findings

1

GPT-4 hurt, not helped.

Adding GPT-4 as a scoring signal decreased F1 by 15 percentage points. Statistical signals were stronger. GPT-4 was retained for plain-English explanations only.

2

The bottleneck was candidate generation, not scoring.

93.5% of missed relationships were never generated as candidates. Every single one had perfect value containment when checked manually. The gap is in naming pattern coverage, not accuracy.

3

Small samples were sufficient.

Sampling 500 values from 238 million rows produced the same accuracy as 1,000 or 2,000. Large samples added noise rather than improving precision.

4

Error analysis revealed a structural naming problem.

A subset of tables use generic column names like entity0 instead of artist_id. These could not be matched by name-based patterns, accounting for the majority of missed relationships.

Media Coverage

Rental Affordability Index

A quarterly report on UK rental affordability, published at Canopy. Cited by over 80 national and regional news outlets including the BBC, Financial Times, Sky News, The Telegraph and Daily Mail.

400+Total citations

80+Unique outlets

National Press · Featured Coverage