| ⏳ 3+ years analytics in production | 📦 13 repos tested, benchmarked, documented | ✅ 1,824 tests 89 to 99% coverage | 🧊 Zero credentials every repo runs on a laptop |
Started where most analysts start: someone needs a number, and the number lives in three places that disagree. Built the reports, then the models underneath them, then the Snowflake and dbt layer underneath those, because the only way to stop the disagreement is to move the definition down the stack until there is exactly one of it. The pivot point came when I noticed the interesting failures were never "the query is slow" but "the query is confidently wrong", and nothing in the BI toolchain was designed to catch that. So I started building the things that do: metric contracts with CI gates, schema classifiers, reconciliation funnels, row level security provers. Now I work on analytics the way a software team works on services, with contracts, tests, benchmarks and an honest account of what breaks.
- Metric governance, where a KPI is defined once and drift from that definition fails the build
- Data quality at ingestion: schema evolution, semantic drift, quarantine and replay
- Snowflake cost attribution, so "which dashboard is expensive" has an answer
- Power BI semantic models under version control, with a diff that explains what a DAX change actually means
| Domain | Stack | Depth |
|---|---|---|
| SQL | CTEs, window functions, query tuning, stored procedures | ██████████ daily driver |
| Dimensional modeling | Star schema, Kimball, SCD Type 2, grain and fan-out | ██████████ daily driver |
| BI and semantic layer | Power BI, DAX, Power Query, RLS, incremental refresh | ██████████ daily driver |
| Warehousing | Snowflake, dbt, DuckDB, sqlglot | ████████░░ production experience |
| Python for data | pandas, numpy, statistical analysis, EDA | ████████░░ production experience |
| Data quality | Contracts, validation, reconciliation, drift detection | ████████░░ production experience |
| Orchestration | Azure Data Factory, Apache Airflow, AWS Glue | ██████░░░░ working proficiency |
| Visualization | Tableau, matplotlib | ██████░░░░ working proficiency |
| Repo | What it shows | The one detail worth clicking for |
|---|---|---|
| snowflake-credit-attributor | Warehouse credits attributed down to individual dashboards | 74% of compute credits were idle, and the idle bucket is separate rather than smeared across queries |
| ops-kpi-anomaly-radar | KPI anomalies attributed to the dimension slice that caused them | A seeded decoy that looks anomalous but is pure mix shift, and the detector correctly refuses to flag it |
| freshness-sla-tracker | Learned per-dataset freshness SLAs instead of one fixed rule | 86% fewer false alarms than "alert if not loaded by 07:00", measured head to head |
| sales-metric-contract-warehouse | Every KPI defined once in a YAML contract, compiled to SQL | The CI gate names the root cause on drift: missing filter, grain fan-out, or wrong measure basis |
| powerbi-semantic-diff | A CI gate for Power BI semantic models | It parses DAX into an expression tree, so it says "ALL was removed, this measure is now slicer-sensitive" instead of showing a text diff |
| nl2sql-guardrail | The safety layer between a text-to-SQL model and the warehouse | 30 of 30 wrong generations caught before execution, and the false refusal rate is published rather than hidden |
The other seven
| Repo | What it shows | The one detail worth clicking for |
|---|---|---|
| scd2-integrity-auditor | 16 structural checks on SCD Type 2 dimensions | Point-in-time referential integrity: facts referencing a dimension version that did not exist yet |
| warehouse-reconciler | Source to warehouse reconciliation in three escalating tiers | Only 21.9% of rows ever reach the expensive row-level diff |
| incremental-refresh-planner | Partition-level refresh planning across a dbt-shaped DAG | It proves the incremental result row-for-row identical to a full rebuild, and caught a case where the naive rule was wrong |
| sql-plan-regression-guard | A CI gate for analytical SQL performance | The noise floor is measured first, so the minimum detectable regression is stated instead of guessed |
| rls-policy-prover | Row level security proved rather than spot checked | It prints the join path that causes each leak, including the small-group disclosure case |
| synthetic-warehouse-forge | Referentially complete synthetic data from a schema plus a profile | Ship a 156 KB profile instead of production data, and the privacy checks are measured, not asserted |
| glue-schema-evolution-framework | Upstream schema changes classified before they land | Semantic drift caught with PSI and a KS test: a column that stayed INT but switched from dollars to cents |
The fastest way to evaluate me is to open any repo above and read the "Hardest problem solved" section.