I'm a BI & Analytics specialist in Vancouver — 4+ years delivering finance-grade (GL/P&L) and operations reporting across sales, finance, and supply chain, most recently at a specialty food distributor. I work the whole path from raw data to the number an executive acts on: pipelines, warehouse models, semantic layers, DAX, and the dashboards on top.
📫 dharma.patel552@gmail.com · 🔗 LinkedIn · 📍 Vancouver, BC · ✅ open to BI Analyst / Analytics Engineer / Decision Support roles
Everything below was built under a single rule: no claim without a number, and no number without a test that fails if it stops being true. Every repo generates its data from a fixed seed, rebuilds end-to-end in GitHub Actions, and re-verifies its own claims on every push — 9,070 automated tests across the thirteen featured repos below. One additional CI-backed utility — the inventory analytics app — brings the public-repo total to fourteen without changing that featured-set test total. A green badge here means it runs, not just that it's written down.
Six of them have hosted endpoints you can click — no local install: retail analytics platform · transaction monitoring · cross-sell console · Ask Your Data · pricing & costing analytics · attribution vs truth. The first is a prerendered static site, so it opens instantly — the others are Streamlit apps that sleep when idle and may take a moment to wake or request a Streamlit sign-in, depending on the host's current visibility setting.
One of these repos used to be a Raspberry Pi voice assistant I built years ago. I rebuilt it into the capstone of this portfolio and kept the git history, because a portfolio that pretends its author sprang fully formed is lying. Start there:
| Project | The one thing to know | Stack |
|---|---|---|
| 💬 Ask Your Data · live ▶ | Grounded text-to-SQL over 71 tables in 11 business domains, asked by voice or text, with two engines. The live demo has no API key and needs none: a deterministic compiler profiles the warehouse (column roles, grains, join graph, a 797-phrase value lexicon read straight out of the data) and binds your words to it, so it answers questions nobody pre-registered. On its 58-question contract it is 46 right, 0 wrong, 12 refused — refusing is the feature. Add a key and Claude writes the SQL instead, through the same read-only guard, structural verifier and default-deny access policy; its prose is then checked back against the returned rows, so a sentence cannot quote a number the query never produced. Six certified metrics carry an owner and a CI-asserted value. Speech runs in-process on open models — faster-whisper and Piper, checksum-pinned — so it listens and answers aloud with no account. I found half its early answers wrong by attacking my own eval four times over, and the README says so. 1,099 tests. | Python, DuckDB, ONNX, Claude API, Streamlit |
| 💵 Pricing & Costing Analytics · live ▶ | The list price is not the price: 18.2% of it never arrives, once every rebate, payment term, freight allowance and return is counted. Margin is expressed on pocket price throughout, because a margin quoted on list is the number that lets a deal look healthy while losing money. The output is not an elasticity, it is a sentence — six ordered actions across 240 products worth 2ドル.65M of annual gross margin, each carrying the reason and a confidence about the evidence rather than the size of the prize. The part worth asking about: the dataset is built so the analysis can be caught being wrong. Elasticity, cost pass-through and segment price sensitivity are seeded deliberately and taken back out through the real code — estimated category elasticities preserve the seeded ordering at rank correlation 0.95, every pass-through estimate attenuated toward zero. Underneath, an SAP-shaped ERP extract with defects injected on purpose: all twelve quality rules find something and the reconciliation balances to 0ドル.00 unexplained. Six marts computed a second time in DuckDB and held to the pandas ones at 1e-8 — which is how the price-band median was caught five percent apart under the same column name. A nineteen-page Power BI project generated from spec — opened in Desktop, refreshed and exported page by page with 0 broken visuals, which is how nine defects no structural test could see were found and fixed, each one now a check in the generator. 2,638 tests. | Python, DuckDB, Power BI (PBIR/TMDL), Streamlit |
| 🔎 Transaction Monitoring · live ▶ | A flat 10,000ドル reporting threshold finds 38 of 60 planted cases and raises 555 alerts doing it. Five explainable rules over an unsupervised model find all 60 on a third fewer alerts, at ×ばつ the precision. The honest half: ablation shows the anomaly model lifts ranking 49% but saves 1.8% at the operating point — so the README says it earns its place on triage order, not on the accept/reject decision. 17 SQL features proven equal to their Python twins on all 100,299 rows. 155 tests. | Python, SQLite, scikit-learn, Streamlit |
| 🏪 Retail Analytics Platform · live ▶ | Sixteen pages that between them printed three different HHI values at the same time, because every page did its own arithmetic. One catalogue now governs 62 definitions: 47 implemented and 15 explicitly unavailable without a named source system. Every implemented metric carries a formula, grain, source table, owner page and gross-or-net basis, pinned by a hand-computed test. The demand planner used to produce a forecast nobody scored; it now leads with WAPE 11.6% and a 40% hit rate from a rolling-origin backtest, and demotes MAPE to a labelled diagnostic because small SKU-month denominators make it least reliable where it looks most alarming. Finance adds statements whose revenue and COGS are read from the sales fact (net margin 2.8%, current ratio ×ばつ, ROA 2.3%); marketing adds CAC 17,667ドル on an 11.0-month payback. Inventory turnover appears twice on two honest bases, and ROMI is withheld outright — spend and wins exist, exposure and a counterfactual do not. Prerendered: 100 charts frozen to SVG, every page verified with networking blocked. 1,412 tests. | Python, Flask, DuckDB, Plotly, Playwright, Docker |
| 🏗️ Supply Chain Control Tower | Medallion pipeline with a three-tier data defense — schema contracts before Bronze, row quarantine with replay, a DQ gate on Gold — each tier proven by CI injecting failures. 10M-row Delta benchmarks, dynamic RLS/OLS verified by impersonated DAX. 659 tests. | Fabric patterns, PySpark, Delta, Power BI |
| 🛒 Customer Recommendation Engine · live ▶ | The fancy two-stage ranker scored 81.2% recall@10; plain collaborative filtering scored 84.9%. The simple model ships, the loss is documented, and CI enforces that the winner keeps winning. The live rep console turns that into what a salesperson actually needs: what to pitch next, why, and what it's worth. FastAPI + Docker serving, A/B framework, 9-page Power BI. 613 tests. | Python, scikit-learn, FastAPI, MLflow |
| 📈 Marketing Attribution & Incrementality · live ▶ | Attribution is the one analytics discipline where everyone argues and nobody can check the answer, because the counterfactual isn't in the data. So I generated one: each user carries a fixed random draw, and a channel's true contribution is measured by re-running that draw with its touches deleted. Six models compete against it and none wins — the exact Shapley value and a Markov chain both lose to a heuristic that fits in a CASE expression. Last-touch hands direct 24.5% of conversions against a true 1.9%. Then a geo holdout settles it: naive pre/post reads 11.0% against a planted 5.5%, difference-in-differences returns 4.0% with an interval that covers it. A second act asks the same questions of a B2B SaaS motion — CRM pipeline, ARR waterfall, NRR vs GRR, quota and capacity, a LookML semantic layer — and finds the segment with the fastest cycle and the best pipeline coverage is the one that loses money on every customer. 91 tests. |
Python, SQL, causal inference, LookML, Streamlit |
| 🧑🤝🧑 HR Attrition Analytics | People analytics with the guardrails real employee data demands: k-anonymity masking, a disparate-impact CI gate (four-fifths rule + Fisher's exact), survival analysis with honest censoring. The flight-risk model uses zero protected attributes — and scores better without them. 629 tests. | T-SQL, Python, lifelines, Power BI |
| 🔄 Supply Chain Analytics (dbt) | dbt Core, staging → marts on dual DuckDB/Snowflake profiles: incremental loads, SCD2 snapshots, MetricFlow semantic layer, Airflow DAG with DagBag validation. 6 unit tests run the model SQL against fixed inputs, so a change to the OTIF threshold or a partly shipped line priced on the wrong quantity fails the build — the data tests would not notice either. 154 dbt data tests across 15 models. | dbt, DuckDB/Snowflake, MetricFlow, Airflow |
| 🏥 Health System Decision Support | Two health systems, one standard. Canadian side: CIHI-DAD-shaped activity (CMG+/RIW, cost per weighted case, ALC, risk-adjusted readmission), SPC with Laney correction — where I found the metric everyone reports is ×ばつ overdispersed and fires 41 signals in 19 of 24 months — and a health-economic evaluation that comes out dominant on one costing perspective and 192ドルk/QALY on the other. Ends in a briefing note and a costed business case. US side: an NRV model pricing 3ドル.6M of open AR at the ~1ドル.7M it will actually collect. Plus Safe Harbor + k-anonymity de-identification with a measured re-identification risk. The Canadian layer now leads the Power BI report — activity and the ALC/flow/SPC page come first, because that is the order a health authority reads them in. No PHI. 604 tests. | Python, Power BI, DAX, SPC, HTA |
| 🧪 Clinical Data Management | A trial database as code: CDASH CRF metadata, an executable Data Validation Specification, SDTM DM/AE/VS with conformance checks, MedDRA/WHODrug coding, and a UAT plan generated from the spec. The generator writes an exhaustive defect manifest — 49 injected, 49 detected, 0 missed, 0 false positives — and that reconciliation caught a real bug where four protocol deviations went silently undetected. The status board is hand-generated SVG, because "stdlib only" is a claim and a chart is not a good enough reason to break it. 255 tests. | Python, CDISC, CDASH/SDTM |
| 💰 GL/P&L Reconciliation | ERP-vs-subledger reconciliation that detects four discrepancy classes and proves every dollar of variance ties to source — then re-runs the same engine, unmodified, over a FOCUS-format cloud bill for FinOps chargeback. 558 tests. | T-SQL, SQLite, Power BI, DAX, Tableau |
| 🚚 Legacy-to-Fabric Migration | SSIS/SSRS → notebook pipeline with parallel-run validation and a GO/NO-GO cutover gate; negative tests prove the validator catches dropped rows, offsetting errors and phantom keys — and, just as importantly, that it doesn't fire on shuffled row order or last-place float drift. Two empty outputs used to return GO; that one is fixed and first in the file. 203 tests. | SSIS, SSRS, T-SQL, PySpark |
The two-stage ranker lost to plain collaborative filtering — documented, and the simple model ships. The gradient-boosted forecaster lost to a moving average — documented, and the moving average ships. The fairness screen fired, and the follow-up analysis showed small-sample noise, not bias — so it became a monitor, not a build-breaker. The health-economics model came out dominant under one costing perspective and not worth funding under the other — so the business case recommends approval explicitly not as a savings initiative, because saying both is what makes the first half believable.
The layered transaction-monitoring detector is the same shape: blending an anomaly model into the score lifts ranking by 49%, and saves 1.8% at the threshold you would actually run. Both numbers are in the README, because only quoting the first one is how a model gets credit for work the rules did.
The biggest loss is the attribution project, where I built a dataset with a known answer and then watched every model fail to find it — including the exact Shapley value and a Markov chain, both beaten by a 40/20/40 heuristic. The reason turned out to be structural rather than fixable: every model that reads only a journey log can see how often a channel was present, never whether it caused anything. That is also where the nicest piece of nuance in the portfolio lives — last-touch is the worst model in the table and the least damaging one for splitting paid budget, because its catastrophic error lands on a channel nobody can buy.
And three bugs found by reconciliation rather than by a person: a visit-window check keyed on the wrong record, which let four protocol deviations through silently; a control chart whose contaminated baseline flagged the stable months instead of the shifted ones; and a p′ chart that rendered completely empty while passing every test it had — the tests checked the maths, and nothing checked that the thing drew. None of them crashed. All three are now tests.
Each repo also has a "things I deliberately didn't build" section: no vector search over business rows where SQL is the right tool (Ask Your Data embeds only schema descriptions), no deep learning on 38 SKUs, no cloud cosplay. Knowing when a technique would be decoration is, I think, the actual skill.
The repos above aren't hypotheticals — they're sharpened versions of problems I've worked professionally:
Two Rivers Specialty Meats — Data Analyst: Operations, Logistics & Enterprise Reporting (Vancouver)
- Defined governed KPIs for P&L/GL and operations (OTIF, inventory turns, days on hand) with Sales, Finance, and Supply Chain; built the Power BI star-schema models and DAX behind them.
- Developed Fabric / ADF / Synapse pipelines (SQL, PySpark) on Bronze→Silver→Gold patterns while owning the legacy SSIS/SSRS estate — the same modernization my migration repo rehearses with a GO/NO-GO parallel-run gate.
- Automated data-quality and reconciliation controls with logging and exception reporting — cutting recurring discrepancies by ~45% and reporting errors by ~30%. FEFO/expiry-risk and lot-traceability analytics for perishable inventory — the direct ancestor of my control tower.
- Secure-by-design delivery: RLS/RBAC, column-level security, masking for cost/margin and people data.
Shivam Investments — Financial Analyst, Data & Reporting (remote)
- GL/P&L reconciliations across sources — the discipline my reconciliation repo turns into a tested engine.
- Automated recurring analysis with Python + SQL feeding Power BI, cutting manual effort ~40%.
BI & semantic modeling — Power BI (DAX, star schema, RLS/OLS, calculation groups, TMDL/PBIR-as-code, VertiPaq tuning), SSRS · Pipelines — Microsoft Fabric, ADF/Synapse, PySpark, Delta Lake, dbt, Airflow, T-SQL, Python (pandas, scikit-learn) — ETL/ELT patterns that carry directly to Talend and equivalent enterprise integration platforms · Health analytics — CIHI DAD/CMG+/RIW, cost per weighted case, ALC and patient flow, risk adjustment by indirect standardisation, SPC (p/u charts, Western Electric, Laney), health economics (ICER, budget impact, PSA/CEAC), CDISC CDASH/SDTM · LLM apps — Claude API, grounded text-to-SQL, eval-driven development · Governance — Kimball modeling, metric dictionaries, data contracts, reconciliation controls, HIPAA Safe Harbor and k-anonymity de-identification, PHI/PII-safe pipeline design · Legacy MSBI — SSIS, SQL Agent, and moving all of it forward without breaking month-end
- 🎓 MPS in Analytics (Applied Machine Intelligence), Northeastern University, Vancouver — GPA 3.76
- 🎓 B.Eng. Computer Science, Gujarat Technological University
- 📜 Microsoft Fabric Analytics Engineer Associate (DP-600) — in progress
- 🏆 First prize at a wildfire-prediction hackathon — Python + Azure + live sensor data (DHT22/LM393), because sometimes the dashboard needs hardware
📚 More certifications & coursework
- Google Data Analytics
- IBM Python for Data Science, AI and Development
- Power BI Data Modelling with DAX · Advanced SQL
- SQL for Data Science: Data Pipelines, DBMS, Data Modeling
If you only have five minutes: open Ask Your Data and type a question of your own — no API key needed. The SQL that produced the number is shown underneath it, along with the table, the measure and the filter it bound, and which words in your question paid for each. Ask it something it cannot compile and it refuses rather than guessing. Then read "The rule: no number without a query" and check the badge is green. Everything else here works the same way.