Open Source
Power BI Dashboard Generator
Nine composable Agent Skills that turn a plain-English request into a valid, branded Power BI Desktop Project: semantic model, visuals and theme included.
PythonAgent SkillsTMDLPBIP / PBIRDAXClaude Code
01 / Work
Featured case study.
ProductionWorldLink US · 2025 – present
A price forecasting platform that ships itself
End-to-end forecasting on Databricks, from ingestion to a registered, promotable model.
- production model versions registered
- 3production model versions registered
- reduction in deployment cycle time
- ~40%reduction in deployment cycle time
- MAPE reduction
- 18%MAPE reduction
- fewer production rollbacks
- ~30%fewer production rollbacks
02 / Open source
Other people's codebases.
databrickslabs/ontobricks
External contributor
A Databricks Labs project that turns Unity Catalog tables into a materialised knowledge graph, with ontology design and reasoning exposed as MCP tools.
Four pull requests to a Databricks Labs repository I do not maintain. Working in someone else's production codebase, with their review standards, their CI and their release process, is the closest thing to a public integration test for how you actually engineer.
PR status last verified 31 July 2026
Open: labelled 'status: in progress', accepted into milestone v0.7.0
A broad question in Graph Chat could freeze the entire app until redeploy: a slow graph read ran directly on the single uvicorn event loop, starving every other request, and was unbounded in both time and result size. This bounds every read server-side, moves the blocking work off the event loop, auto-sizes the worker pool to the instance, and degrades a slow query to a clean per-request cancellation instead of a global stall.
Read how I debugged itOpen: awaiting review
Registry configuration resolved in the wrong precedence order, so a stale per-session value could shadow the configured environment and send binary archives and Delta views to the wrong Unity Catalog Volume. A single-file precedence fix with regression coverage.
Closed unmerged: taken over by a maintainer
Knowledge-graph sync aborted whenever a mapped literal exceeded Postgres' 2704-byte B-tree limit. Proposed re-keying the companion tables on a generated SHA-256 column and making the object-bearing indexes size-guarded partial indexes, with a best-effort migration for existing tables.
Closed unmerged: fix had already landed upstream
I hit a 403 filtering a locked graph version and traced it to an over-broad prefix match in the permission middleware. The fix was already in upstream, so I offered a middleware dispatch test to stop the regression returning rather than a duplicate source change.
03 / Projects
Things I built.
Personal
SEC Financial Chatbot
Ask questions about US public-company filings and get answers with citations back to sec.gov. Pulls live XBRL facts and filing sections, with no backend and no database.
ReactTypeScriptSEC EDGAR XBRLOpenRouterBYOK LLM
Personal
Stock Market ETL Pipeline
End-to-end medallion pipeline on Databricks taking raw market data through Bronze, Silver and Gold into a dashboard-ready analytics layer.
PythonDatabricksDelta LakePySparkMedallion
Personal
OCR Text Vision Pro
Vision-model document extraction that runs entirely in the browser: upload an image or PDF, get structured text back, then interrogate it in chat.
ReactTypeScriptVision LLMsOpenRouterpdf.jsStreamlit
Personal
LinkedIn Data Visualizer
Drop in your LinkedIn export ZIP and get an interactive dashboard of your network and activity. Parsing happens client-side, so the data never leaves your machine.
ReactTypeScriptClient-side parsingOpenRouterHTML report
04 / Writing
Thinking out loud.
Explaining a system is the fastest way to find out whether you actually understand it. I write at The Brian Journal about data engineering, AI and the things that surprised me.
- Databricks·20 min read
Databricks Data + AI Summit 2026: The Lakehouse Becomes the Agentic Control Plane
A technical deep dive into Genie One, Unity AI Gateway, Lakeflow, LTAP, Lakebase and Lakewatch, and what each announcement actually changes for data and AI engineers.
Read - Book review·13 min read
Principles and Patterns of Building AI Agents: An Honest Review of Mastra's Two-Book Series
Principles teaches what to build; Patterns teaches how to keep it alive in production. A grounded look at where each one earns its place and where it doesn't.
Read - Book review·10 min read
30 Agents Every AI Engineer Must Build: An Honest Review After Three Months
Why Imran Ahmad's book is really a pattern library for production agent engineering rather than a catalogue of thirty demos, and who should actually read it.
Read
05 / Experience
Where I’ve worked.
- Sept 2025 – Present
AI Data Engineer @ WorldLink US
- Architected and deployed an end-to-end price forecasting platform on Databricks using MLflow, XGBoost and CI/CD, registering three production model versions and cutting deployment cycle time by roughly 40%.
- Built a multi-model inference service (React + Databricks Apps) serving five integration patterns, holding sub-two-second response times under a 1,000-concurrent-user load test.
- Implemented automated MLOps workflows (model validation, versioning and Dev→Prod promotion), reducing production rollbacks by about 30% and release errors by about 45%.
- Improved forecast accuracy through feature engineering and hyperparameter tuning, achieving an 18% reduction in MAPE against the baseline time-series model.
- Engineered an LLM-powered natural-language-to-BI pipeline integrating Databricks Genie with Power BI, cutting manual report creation time by around 60%.
- Automated SQL-to-visual translation and YAML ↔ TMDL metadata conversion with Agent Skills and LLM orchestration, reaching 92% query-to-visual accuracy on the internal evaluation set.
- Aug 2024 – Jan 2025
Graduate Teaching Assistant @ George Mason University
- Led weekly sessions for 40+ students on optimisation modelling in Excel with Frontline Solver, improving accuracy on linear and mixed-integer problems by around 30%.
- Ran interactive labs on Gurobi in Python, raising measured student engagement scores by about 20%.
- Jan 2022 – Aug 2023
Senior Product Analyst @ Media.Net Software Services Pvt. Ltd.
- Partnered with engineering, product and campaign teams to ship data-driven changes to AdTech serving performance.
- Ran 10+ concurrent weekly A/B tests on creative templates using in-house tooling, improving recommendation accuracy by around 50% and campaign ROI by around 70%.
- Automated recurring analysis in Python and SQL, removing roughly half the manual effort from the weekly reporting cycle.
- Led execution across 5+ initiatives in Jira and Confluence, delivering all of them on schedule.
- Mentored 10 junior analysts on ETL design, SQL and reporting practice.
- July 2019 – Dec 2021
Product Analyst @ Media.Net Software Services Pvt. Ltd.
- Built stored procedures automating ingestion and transformation of 5+ GB of daily ad-serving data into SQL Server, halving batch processing time.
- Integrated new ad platforms over REST APIs alongside engineering, increasing impression delivery by around 20%.
- Delivered weekly campaign performance and optimisation reporting to 10+ clients in Tableau and Excel.
06 / About
Who I am.
I build the unglamorous parts of machine learning: the ingestion that does not silently drop rows, the model registry that makes a rollback a one-line change, the inference layer that stays up when six different consumers hit it at once.
I started in AdTech at Media.Net, where I spent four years as a Product Analyst and then Senior Product Analyst turning ad-serving data into decisions people actually made. That taught me the thing no course does: a pipeline is only as good as the question it answers. I moved to the US for an MS in Data Analytics Engineering at George Mason, taught optimisation modelling as a graduate TA, and now work as an AI Data Engineer at WorldLink US on forecasting and LLM systems running on Databricks.
Lately most of my time goes to the seam between LLMs and analytics: getting a natural-language question to become a governed SQL query, then a semantic model, then a dashboard someone trusts. I also contribute to Databricks Labs' OntoBricks, where debugging an event-loop stall in someone else's production codebase taught me more than any tutorial has.
I write about all of it at The Brian Journal, mostly because explaining a system is the fastest way to find out whether I actually understand it.
Beyond the screen
- Die-hard Real Madrid supporter. The Ronaldo argument is settled, thanks.
- Hiking and trail walks, pretending I'm Bear Grylls for an afternoon.
- Folding origami animals when I need to think without a screen.
- Collecting productivity hacks, then over-engineering them into side projects.
Databricks & Lakehouse
Daily, in production, since 2025
PySparkDelta LakeUnity CatalogMedallion ArchitectureDatabricks JobsDatabricks AppsModel ServingGenie
ML & MLOps
Production forecasting platform, 3 registered model versions
MLflowXGBoostscikit-learnTime-series forecastingFeature engineeringHyperparameter tuningModel registry & promotionCI/CD for models
LLM & Agent Engineering
Shipped agent skill toolkits and retrieval apps in the open
RAGPrompt engineeringAgent SkillsMCPLangChainOpenRouterHuggingFaceClaude Code
Data Engineering
Five years building ingestion and transformation pipelines
SQLPythonETL / ELTAirflowdbtSnowflakeSQL ServerREST API integration
BI & Semantic Modelling
Automated the SQL-to-visual path end to end
Power BITMDLPBIP / PBIRDAXTableauSemantic modelsSelf-service enablement
Platform & Tooling
How the work actually ships
Git & GitHub ActionsDockerTypeScriptReactGoogle CloudMicrosoft FabricPytest
07 / Credentials
Verified, not just claimed.
Aug 2023 – May 2025
Master of Science, Data Analytics Engineering
Concentration: Data Modeling/Warehousing and Database Administration
George Mason University, College of Engineering and Computing
Verify this degreeA 30-credit multidisciplinary programme combining statistics, computer science and operations research, aimed at the engineering side of analytics rather than reporting.
Relevant coursework
Statistical MethodsData MiningMachine LearningBig Data AnalyticsOptimisation & Simulation
July 2015 – June 2019
Bachelor of Engineering, Information Technology
University of Mumbai, Don Bosco Institute of Technology
Four-year engineering programme covering the fundamentals I still use daily: databases, operating systems, networks and software engineering.
Relevant coursework
Database Management SystemsOperating SystemsComputer NetworksSoftware EngineeringWeb Technologies
7 certifications, each independently verifiable
08 / Contact
Let’s talk.
Have a question, a role, or a system you want a second opinion on? The fastest way to reach me is email; I reply to everything that isn’t automated.
Open to AI/ML Data Engineer, Analytics Engineering and Data Platform roles. Based in Dallas, TX; open to relocation and remote.
- briancastelino07@gmail.com
- Dallas, TX, USA