The analyst half decides what is worth measuring. The engineer half builds the thing that measures it — governed Databricks systems whose numbers are published and enforced, including the ones that came out badly.
Technical portfolio → · LinkedIn
Testing platform claims before building on them. Can one Databricks SQL function replace a
tuned forecasting pipeline? databricks-forecast-planning-cockpit answers that
under a protocol frozen before the test year was downloaded — and publishes the part of
the answer that came out badly. Released as
v1.0.0.
Structure-aware, multilingual PII reduction for Databricks · released
v0.1.0· Python · Presidio + spaCy
Reduces PII inside free-text columns an operator names. A ticket id survives, a timestamp and a speaker label survive; the name and the email do not. English, German and Greek. Every published number is a regression gate — 56 of them, across three corpora and both provider chains — so no figure in the documentation can move without a test failing.
| Executed | driver-path parity on a real Azure Databricks workspace; the service hosted as a Databricks App and driven over HTTPS |
| Not executed — and it says so | the distributed mapInPandas path is shipped and has never run; ADDRESS is in the taxonomy and nothing detects it; Greek PERSON recall is published as 0.500 rather than rounded up |
| Not claimed | it is not an estate scanner, and it promises no compliance outcome and no guaranteed anonymization |
Inspect · what was actually executed · 36 decision records · the measured baseline · providers and their limits
Frozen-protocol forecasting benchmark · released
v1.0.0· Databricksai_forecast· Prophet · Delta · Asset Bundles · AI/BI
ai_forecast v2 against a tuned Prophet pipeline and a weekly seasonal baseline on 21 NYC
311 daily-demand series, 28 days ahead from 12 monthly origins across 2025. The SQL
function won on accuracy — median series error 17.1% against 19.5%, best on 17 of 21
series — with no training code and no tuning. It also ran 14.3% low, below actual
demand at every origin, so under the pass rule written before 2025 was downloaded it does
not qualify as a replacement.
| Executed | all 756 forecast cells scored with zero failures, v2 as fresh inference on a Databricks trial workspace; predictions stored in Delta and served by a code-managed AI/BI dashboard; 85 tests in CI |
| Not like-for-like | v2 ran on Databricks while Prophet and the baseline ran locally, so runtimes are reported but not compared as a speed test |
| Not claimed | a retrospective backtest on one public dataset — no universal winner, and no staffing, capacity or savings result |
Inspect · benchmark protocol · evidence index · what v2 needed to run
Released Databricks accelerator ·
v1.0.0· analytics engineering · forecasting · AI/BI · Genie
Answers an operating question rather than a modelling one: when is a new-hire cohort becoming production-ready, and what evidence supports the decision? Four governed source tables become learning curves, volume-aware diagnostics, forecasts, process-control views, an AI/BI dashboard, a Genie space and an optional evidence-grounded action workflow. Release quality gates run in CI.
Inspect · validation record · architecture
Released reference ·
v0.1.0· governed retrieval · retrieval evaluation
Built around one testable claim: fixed-window retrieval loses document structure and evidence relationships. It preserves the hierarchy and expands exact one-hop CFR references as separately citable evidence, against a date-pinned public eCFR corpus with a committed evaluation set and live Databricks evidence.
Inspect · evaluation design · retrieval design
| prophet-forecasting-mlops | A compact, reproducible batch-forecasting reference. Forecasting behaviour stays in testable Python; Databricks-specific code is confined to delivery, tracking and persistence. A seeded synthetic source makes the contracts reviewable without private data — and the recorded run counts are execution and contract checks, not accuracy claims. |
| databricks-genie-deployment-kit | Semantic analytics managed as code: room configuration, semantic metadata, SQL examples, benchmark questions, deployment scripts and operating playbooks as reviewable assets, with a public-data Olist example. Durable repository-native dashboard evidence is still pending, because the published dashboard is not anonymously accessible. |
| speechanalytics-databricks-pipeline | A 16-stage contract-first speech-analytics design with per-call failure isolation and guards against raw transcript text reaching analytical outputs. No recorded successful Databricks pipeline execution, so it stays labelled a prototype. |
- Gates before claims. A published number is enforced by a regression test, not restated from a notebook run nobody can repeat.
- Evaluation kept out of the pipeline. Ground truth comes from a generation manifest, so it is derived rather than reverse-engineered after the fact.
- Reproducibility as the default. Seeded synthetic data, pinned upstream revisions, locked environments, public-safe fixtures, CI on every push.
- Limitations written down. Each repository states what has not been executed and what it does not claim. In someone else's work, that is the list I read first.
Three learning-stage repositories stay public because they show how the work developed, not because they stand beside the systems above — an early Spark ML notebook comparing loan-default classification approaches, and two absenteeism studies covering feature engineering and workforce modelling and model selection across departments. They have notebook outputs and written findings; they do not have the reproducible environments, tests, CI, deployment boundaries or public validation data the newer work does.
Technical portfolio — the full map, with case studies and where to inspect each proof · LinkedIn