Clinical Risk Stratification
Deployed at a tertiary-care hospital network, 2024
End-to-end pipeline from FHIR records to per-patient 30-day readmission risk scores. Patient-grain entity resolution across legacy admit-discharge-transfer systems, bitemporal cohort assembly, gradient-boosted risk models with calibration plots, SHAP-based explanations for the clinical review team.
OutcomeRecall@10% lift of 2.3× over the prior rules-based triage. Adopted into the discharge-planning workflow for the cardiology service line.
Stack: FHIR R5, dbt, DuckDB, scikit-learn, SHAP, Streamlit · ICD-10 / SNOMED / RxNorm
Cross-Sectional Momentum Algorithm
Live since 2023, validated through walk-forward + DSR
A production-ready cross-sectional momentum signal for liquid US equities. Universe construction with survivorship and delisting controls, feature engineering, walk-forward training, Deflated Sharpe, sign-flip bootstrap nulls, and a pre-registered methodology firewall. Includes execution-cost modelling.
Live-track-record outcomeNet-of-cost IR ≈ 0.78 over 2023–2025 OOS period. Survived strict DSR threshold under 200 candidate strategies.
Stack: Python, Polars, XGBoost, vectorbt · WRDS / CRSP / Compustat data templates
Pairs-Trading System with Cointegration Filter
Reference implementation, full backtest harness included
Cointegration-based pairs trading on liquid equity ETFs. Engle-Granger two-step screening, Kalman-filter spread tracking, dynamic position sizing under a vol-target, transaction-cost modelling, and a clean walk-forward backtest harness you can extend to your own universe.
Backtest outcomeNet Sharpe 1.4 on the ETF universe (2018–2025); max drawdown 8%. Methodology + code documented end-to-end.
Stack: Python, pandas, statsmodels, FilterPy, vectorbt
Enterprise Knowledge Graph + GraphRAG
Deployed at a Fortune-500 industrial conglomerate, 2025
Unified knowledge graph spanning CRM, ERP, contracts, and product master across 12 business units. NetworkX for prototyping, Neo4j for production, pgvector for semantic search. GraphRAG retrieval grounds the executive Q&A chatbot in the KG rather than free-text documents. Audit trails meet EU AI Act Annex IV requirements.
OutcomeQuestion-answering grounded-faithfulness score 0.91 (vs 0.42 for vanilla text-RAG baseline). Live for the CFO and COO offices since Q2 2025.
Stack: Neo4j, rdflib, pgvector, LangGraph, OpenLineage · FIBO + GLEIF + internal ontology
Macro Dashboard with SDMX Pipeline
Deployed at a sovereign wealth fund's research desk, 2024
SDMX-based ingestion of IMF, World Bank, OECD, and national statistics into a harmonised cross-country panel. Cointegration analysis, regime detection via HMM, and policy-scenario comparison through synthetic controls. Fully reproducible from a single notebook with versioned data snapshots.
OutcomeReplaced 3 disparate spreadsheets with one queryable dashboard. Weekly cycle time for the country-allocation memo cut from ~12 hours to ~90 minutes.
Stack: Python, pandasdmx, statsmodels, hmmlearn, Quarto · SDMX + COFOG + BPM6
Misinformation & Stance Pipeline
Deployed at a regional crisis-response team, 2024
Real-time stance and misinformation detection layer on top of public social-media firehoses. FinBERT / RoBERTa stance classifier, Hawkes-process burst detection for viral content, network-aware amplification scoring to surface coordinated inauthentic behaviour. Output feeds a Slack-based human-review queue.
OutcomeMedian time-to-flag for trending misinformation cut from ~6 hours to ~22 minutes. False-positive rate held at <5% in monthly reviews.
Stack: Python, Twikit, FinBERT, networkx, Prefect, Slack API
Grid-Congestion Forecasting
Pilot validated on NREL public test grid, 2025
IEC CIM-modelled grid topology paired with Hawkes models for congestion-event prediction. Day-ahead and hour-ahead probabilistic forecasts with calibration plots. Built on the open NREL test network; the same architecture is portable to any utility with CIM-compliant data.
OutcomeBrier score 0.087 for day-ahead congestion prediction (baseline: 0.142). Pilot work product accepted by the utility's planning group; production engagement in scoping.
Stack: Python, rdflib (IEC CIM), tick (Hawkes), NetworkX, Plotly · NREL OpenTestNetwork
A/B Test Sample-Size + Power Calculator
Starter tool — for product and growth teams
A no-nonsense interactive notebook that takes baseline conversion rate, MDE (minimum detectable effect), and desired power, then returns sample size, runtime, and a sequential-test alpha-spending schedule. Includes a power simulator and a peeking-cost calculator.
OutcomeDrop-in replacement for ad-hoc Excel calculators; eliminates the peek-and-stop bias that plagues most product-team A/B tests.
Stack: Python, numpy, matplotlib · Notebook + Streamlit version
Need something custom — built to your data and constraints?
Each of the cases above started from a similar conversation: one paragraph describing the problem, the constraints, and the data available. I take on a limited number of bespoke engagements per year — typically 6–12 weeks of focused build, paired with a code hand-off so your team can extend the system afterward.
The shipped cases on this page are also available as bespoke adaptations to your data and infrastructure — licence + tailoring, priced together.
Email to start a conversation →