Jakub Sobolewski

• Assessing the technical quality of Appsilon projects to highlight risks that may impact projects’ success. Collection of metrics for C-level reporting.
• Running a quarterly technical-health review across the 60-person Appsilon tech team.
• Refined the process from human-driven to self-service via a more detailed questionnaire and a dedicated app (React + CopilotKit + FastAPI + PydanticAI).
• The app implements:
  • Automatic AI-assisted draft reviews of surveys including projects’ historical data.
  • AI-assistant for intra-project and cross-project reviews.
  • MCP server (FastMCP) to interact with app data via external chat apps (Claude Desktop). MCP server interacts with the app data via a REST API.
  • Slack integration for scheduled survey reminders and survey completion notifications.
• Tested with pytest (backend) and Playwright + MSWJS (frontend).

Led development of multiple AI-enabled applications with a team of software and data engineers.

• AI-powered pharmacovigilance platform (React + FastAPI). Surfaces issues across case reports for manufactured drugs. AI-assisted chat interface for querying and summarizing data across reports. Agentic code built with PydanticAI. AI-agent integrated with PostgreSQL and MongoDB. AI-agent evaluation (Pydantic Evals) consists of a golden set collected from business experts, evaluation report is created on schedule on production dataset. Per user feedback, compiling data together now takes hours instead of weeks. Tested with pytest (backend) and Playwright + MSWJS (frontend).

• Clinical data quality platform (React + FastAPI). Monitors data quality across 2 clinical studies with custom ECharts visualizations. Combines SQL-based validation with AI-assisted analysis to identify potential data issues (binary classification), support review workflows, and enable natural-language querying of study data. AI-agent evaluation (Pydantic Evals) consists of verifying whether AI classification complies to human classification of potential data issues. AI-agents integrated with PostgreSQL and MongoDB. Built agentic backend components with PydanticAI. Comprehensive backend test suites using pytest (backend) and Playwright + MSWJS (frontend).

• Internal analytics portal (React + FastAPI). Developed internal applications for discovering analytics products and reports and for managing content submissions, improving visibility of analytics teams’ work across the organization.

• Clinical data quality monitoring app (R/Shiny). Built an interactive dashboard for monitoring clinical data quality metrics (3 studies) with Apache ECharts visualizations. Developed the application outside-in using Behavior-Driven Development (BDD) with Cucumber acceptance tests. The application has been used as part of the team’s (~10 people) regular data review process for over two years with no reported production defects since launch.

• JazzAIR Shiny apps framework. Took over the framework from an internal team. Delivered a framework that ~15 developers across 6 clinical studies use to build Shiny apps for analysis of ADaM clinical datasets; ~20 apps built on it to date. Improvements I drove with the team cut the time to deliver an analytics module/page from ~2 weeks to ~2 days.
• Worked with business experts and developers (users) on the direction of the framework.
• Presented the test automation and development approach at PHUSE EU Connect 2025 Hamburg

• Built a framework for defining and implementing modeling Scenarios, driven by automated testing (unit, module, acceptance) to verify business logic.
• Made the modeling Scenarios and their business logic verifiable with Cucumber acceptance tests whose outcomes are tables showing how each Scenario affects the data — so business experts can see the logic works exactly as intended. Previously, correctness wasn’t guaranteed and had to be checked by hand through the app UI.
• Replaced manual Excel export/reupload with persistent in-app session storage using Pins on Posit Connect, saving time previously lost to external download/upload for state management; improved data-retrieval speed with Pins and DuckDB.
• Refactored a complex health scenario-modeling Shiny app for an Australian government health agency from {golem} to {rhino}, making it enterprise-ready.
• Modularized the codebase, set up continuous integration, refactored the UI/UX for database views and Scenario creation, and delivered personalized R/Shiny training to the client team.
• Case study: https://www.appsilon.com/case-studies/improving-health-scenario-modeling-with-rhino

• Developed R/Shiny apps for:
  • visualizing clinical trial data,
  • visualizing molecular data,
  • visualizing data in translational research.
• Highlight: sped up data loading from 7 min to 2 min.
• Case study: https://www.appsilon.com/case-studies/critical-apps-wait-time-from-7-minutes-to-30-seconds

• Performance analysis of the genomics app; proposed which parts to speed up and how; prepared a list of practices to improve codebase quality.
• Highlight: reduced initial loading time from 2 min to 45 sec.

• Implemented a provided design into an existing app.
• Development of a custom, map-based visualization of geospatial data with Leaflet.

• Gathered socio-economic data for building a model to help understand hate crimes in the US.
• Highlight: built an easily extendable {targets} pipeline producing a single ML-ready dataset.

• Consulting on best practices for working with databases in R.

• Built a POC from a provided dataset with specific visualization and control requirements.
• Custom visualizations in ECharts.
• Highlight: the anonymized app was presented to another client and led to a contract.