Sonja Bacinski

My career started in finance. In 2018, as a finance intern in Denmark, I prepared the monthly bank reconciliations: the company's records on one side, the bank's on the other, and a reason for every difference. I didn't know it then, but that job would follow me into engineering.

Later I co-founded a small e-commerce business, trained in Salesforce and joined Globant's graduate programme as a junior Salesforce consultant. Alongside all of it I studied computing at Birkbeck, University of London: a foundation degree in 2024, then a BSc in 2026, which I finished while working full-time.

That job was at Exaloan, a fintech connecting institutional investors with loan originators. In my first year I built a pipeline that went into production. It reads the statements partners send by email, files them, classifies every transaction and reconciles the monthly cash balances: the same kind of work I used to do by hand. For my thesis, I then spent six months modelling the company's lending data as a graph.

Now I'm building Liquet in public. Rules match a firm's cash records against its bank statement, an AI agent explains what's left, and a person signs off.

I'm looking for a role where I connect systems and work closely with the people who use them: integration, implementation or data engineering in financial services. My favourite moment in any project is when a messy process becomes something people trust.

• Python
• SQL
• ETL & data pipelines
• REST APIs & OAuth2
• Data mapping & validation
• Reconciliation
• ArangoDB (graph & document)
• AWS Lambda
• FIX 4.4
• Git

“What stood out most was her growth… Her communication was a real strength throughout—clear, proactive, and dependable.”

• Email-Ingestion & Reconciliation Pipeline

  Statement emails filed, classified and reconciled automatically. In production at Exaloan.

  Details

  Challenge: Investment-account statements arrived by email as PDFs and spreadsheets, each partner with its own layout. Someone had to download, rename, file, read and check every one by hand.

  Solution: I built the pipeline that does all of it as soon as an email arrives. It renames each file from its content, files it to SharePoint, classifies every transaction and reconciles the monthly cash balance against the partner's own figures. Anything it can't identify waits for a person instead of being guessed.

  In production at Exaloan · Code private

  • Python
  • pandas
  • pdfplumber
  • Microsoft Graph API
  • AWS Lambda
  • ArangoDB

• FIX Allocation Transformer

  FIX 4.4 allocation messages flattened into one CSV row per account.

  Details

  Challenge: When a broker splits a block trade across client accounts, the split arrives as a FIX 4.4 AllocationInstruction: dense, nested repeating groups that are hard to check or use directly.

  Solution: A Python tool that parses these messages and flattens them into one CSV row per account. It resolves each party by its role, including give-ups, rejects messages whose group counts don't add up, and doesn't duplicate a message that's processed twice. Built test-first on synthetic data with pytest; gap analysis and a command line are next.

  Open source · In progress

  • Python
  • FIX 4.4
  • pytest

• Liquet

  Month-end reconciliation: rules, an AI agent and a person signing off.

  Details

  Challenge: Every month a firm checks its cash records against its bank statement. Most lines match; the rest have to be explained by hand, and an auditor has to be able to follow every decision.

  Solution: A reconciliation system where rules match what they can, an AI agent investigates what's left through a small set of tested tools, and a person approves each explanation. When the evidence isn't enough, the answer is non liquet: it isn't clear, so a person decides. Built so far: a statement reader with self-checks (totals, running balance, dates) and tests. Next: the matching rules.

  Read the article: Model the Job, Not the Outcome

  Building in public · Open source · Synthetic data

  • Python
  • LLM agents
  • Graph modelling
  • Reconciliation

• Graph ETL Pipeline | BSc Thesis

  Loan data loaded into ArangoDB as a graph: about 58,000 nodes and 110,000 edges.

  Details

  Challenge: In a lending business, loans, borrowers and investors are tightly connected, but a document store doesn't keep those connections directly queryable. Asking "who is connected to what" means rebuilding the links through repeated lookups.

  Solution: I worked with developers and finance colleagues to understand the data, designed a graph schema around how an investment moves through its lifecycle, and built a Python pipeline that loads it into ArangoDB: about 58,000 nodes and 110,000 edges, zero orphan edges, safe to re-run. The thesis then tested what the graph actually buys you, on six real questions.

  Read the article: Does the Graph Earn Its Cost?

  Prototype · Code private, built with my employer

  • Python
  • ArangoDB
  • Graph modelling
  • Raphtory

• Receivables system for a small business

  144 customers and 73 Excel templates became one app in daily use.

  Details

  Challenge: A family business invoiced 144 customers from 73 separate Excel templates, so tracking invoices, payments and balances was slow and easy to get wrong.

  Solution: I first brought everything into one Excel system, then replaced it with an app the business now runs on every day: invoices, cancellations, credit notes, payments, a ledger for each customer and quarterly rebates. I designed the database before writing any import code, so the old templates came across into one consistent model.

  In daily use · Code private (family business)

  • Python
  • FastAPI
  • SQLite
  • Data migration
  • Accounts receivable

The same five habits run through everything I build.

I start from the person doing the job. I listen and ask questions until I understand the process completely. Before I designed Liquet’s graph, I followed the person closing the month, question by question, and their questions became the design. At Exaloan, I spent almost as much time investigating as building.

I'd rather stop than guess. I work this way myself: when I don’t understand something, I ask instead of guessing, and I build my systems to do the same. A silly question at the start is cheaper than a wrong system at the end. My pipeline leaves any file it can’t identify for a person instead of forcing it into the database. Liquet has a verdict for the same idea, non liquet: when the evidence isn’t enough, a person decides. In finance, a quiet wrong answer costs more than an open question.

I build in small, tested steps. I like to see things working quickly, but in finance a fast wrong answer costs more than a slow right one. So I build one piece of logic at a time and test it before I move on: the normal case, the edge cases, and the bad input that should be rejected rather than quietly accepted. Liquet's statement reader was built this way, with pytest.

I go looking for new ideas, then test them. I love inventing, and I never run out of ideas. Graph databases started as my thesis topic and opened up a whole new world for me. I went to Connected Data London for the conference and masterclass, and I met people who build these systems, to learn how they think. Liquet is my next experiment: I believe a graph will help its agent find answers, but I want to see it working with my own eyes before I claim it. For me, an interesting idea is the beginning of the investigation, not the conclusion.

I build trust with the people I work with. At Exaloan, I worked remotely with a team in Germany, so I made time for small check-ins that weren’t about the task, while keeping it professional. I deliver what I promise, and I only promise what’s realistic. I prefer to demonstrate what an idea can do before making promises about it. When I build a tool, I train the people who will use it in the way they learn best, with a recording or written steps. Work goes better when people trust you, and so does asking for help.

• Does the Graph Earn Its Cost?

  1 October 2026

  I built a graph next to a real document store and asked both the same six questions. The graph didn't reveal anything new: it moved the work of connecting records into one tested pipeline.

• Model the Job, Not the Outcome

  9 October 2026

  How I designed Liquet's graph by following the person closing the month, question by question, and found three gaps before writing any graph code.

• BSc Computing

  Birkbeck, University of London · 2024–2026

  Studied while working full-time at Exaloan. Thesis: Transforming Loan-Related Financial Data from NoSQL Collections into a Graph Model Using ArangoDB, developed with Exaloan. Read the article →

• Foundation Degree in Computing

  Birkbeck, University of London · 2024

  Web development, databases and object-oriented programming. Thesis project: PATHFINDER, a Salesforce learning platform.

  Details

  Focused on web development, databases, and object-oriented programming (Java, Python, PHP, SQL, HTML, CSS, JS). Thesis project, PATHFINDER, repurposed Salesforce as a learning platform, a data model of 9 custom objects with an Experience Cloud student portal and an interactive quiz app built in Lightning Web Components and Apex, featuring difficulty levels and progress tracking.

• Foundation Degree in Financial Management

  International Business Academy, Kolding, Denmark · 2018

  Finance coursework, and a thesis with Aperian Global on turning unused behavioural data into a product.

  Details

  Coursework: Financial Methodology, Advanced Statistics, Financial Market and Advising, and Financial Control. Thesis, Transforming with Big Data, written with fellow student Brendan Hutchinson in collaboration with Aperian Global, a global cross-cultural leadership consultancy. We identified that years of behavioural data from the firm's cultural-training tools sat unused as a stored cost rather than a revenue driver, and proposed turning it into a product: a phased plan using data mining and machine-learning classification to surface patterns in that data and generate predictive sales-and-client profiles, grounded in a full strategic and financial analysis of the business.

• Finalist, Pioneer 1.0 programme, Birkbeck: pitched Zolibri 2022
• Connected Data London: conference and masterclass 2025
• Salesforce Pathfinder Programme | Administrator & Developer Track 2022

If you're hiring for integration, implementation, data engineering or automation roles in financial services, or you work in fintech and want to talk, I'd be glad to hear from you. I'm based in London, with full right to work in the UK.