Work

Context

SaaS leadership teams typically have KPI dashboards, product analytics, and churn tooling scattered across separate systems, with no single place to see health, risk, and growth signals together, and no way to just ask the data a plain-language question.

My Role

Sole designer and builder, from data model and ML services through the full-stack application and deployment.

Actions

• Built a full-stack platform (Encore.ts backend owning its own Postgres, Next.js and React frontend) shipping 6 modules: executive KPIs, product analytics, customer health scoring, ML-driven segmentation, churn prediction, and an AI analyst copilot.
• Trained and deployed real ML models in a dedicated Python service rather than mocked approximations: k-means clustering into 4 customer personas and an XGBoost churn classifier validated on a held-out train/test split.
• Built a Gemini function-calling AI copilot that answers business questions by calling the platform’s own live endpoints as tools, grounding every answer in real data instead of fabricating numbers.

Outcome

A fully deployed, three-service product (Vercel, Encore Cloud, Railway) that turns raw subscription data into segmentation, churn risk, and conversational insight in one place.

Context

July 4, 2026 — America’s 250th anniversary — was projected to be a record travel window, but no public tool forecast its daily demand or showed where impact would concentrate using only free, cited data.

My Role

Sole analyst and builder, from data pipeline through model selection, dashboard, and live validation.

Actions

• Built an ingestion pipeline over TSA, EIA, weather, and holiday data, then selected Prophet with a COVID-period indicator over SARIMA on a 2024–2025 holdout (6.33% vs 9.22% MAPE).
• Validated the forecast against reality as TSA published actuals: all 9 days of the July 4th window landed inside the uncertainty interval.
• Added a City Impact layer scoring 8 cities on exposure to anniversary demand — air capacity, events, and demographics — with live-adjustable weights.

Outcome

A deployed dashboard whose forecast held up against real-world actuals — 4.5% live error, beating its own 6.33% holdout benchmark.

Context

Claims analysts need to know why a claim was denied and which policy rule applies — answers buried in long policy documents and easy for a generic chatbot to fabricate.

My Role

Designed and built the full retrieval-augmented application, frontend to backend.

Actions

• Built an assistant that answers denial questions grounded in the actual policy documents and cites the governing rule id (e.g. COV-002.2).
• Implemented retrieval with Postgres full-text search over chunked policy docs — no external embeddings service — and streamed answers via Server-Sent Events.
• Re-architected a Streamlit prototype into a Next.js (Vercel) and Encore TypeScript (Encore Cloud) monorepo driven by GitHub CI/CD.

Outcome

Traceable, policy-grounded answers — including drafted appeals — that cite their source rule rather than guessing.

Context

A compliance process in the hazardous-materials industry was losing efficiency to unclear KPI drop-off and hidden workflow bottlenecks.

My Role

Lead analyst on the capstone team, owning problem framing through recommendation.

Actions

• Mapped the end-to-end process and built a multi-stage funnel to locate where KPIs dropped off.
• Ran hypothesis-driven analysis to isolate the bottlenecks with the largest impact.
• Translated findings into prioritized, defensible recommendations for stakeholders.

Outcome

Recommendations improved tracking efficiency by roughly 30 percent.

Context

Procurement and operations leaders across five business units lacked a single, trusted view of performance.

My Role

Business analyst owning the reporting solution end to end.

Actions

• Defined and consolidated 15+ KPIs spanning procurement and operations.
• Built a Tableau and Excel reporting suite tailored to senior stakeholders.
• Standardized metric definitions to cut reporting discrepancies and rework.

Outcome

Faster, better-aligned decisions across all five business units.

Context

Early-stage founders had no single screen to answer the questions that decide a startup’s survival: how long until cash runs out, where the money is going, and whether anything is going wrong.

My Role

Sole designer and builder, from data model through the analytics engine and UI.

Actions

• Built a pure, fully unit-tested finance engine for cash runway, burn rate, revenue forecasts, and vendor concentration risk via the HHI index.
• Added anomaly detection that automatically flags duplicate vendor payments and spend spikes without being told where to look.
• Layered a deterministic rule-based analyst that ranks every finding by severity and explains, in plain English, what changed and why.

Outcome

A single decision screen that turns raw financial data into ranked, plain-English actions — runway, risk, and anomalies at a glance.

Context

Customer-lifetime-value models are usually scored on the whole population, which lets a model look accurate just by re-identifying already high-spend customers while missing those whose spend is about to climb.

My Role

Author of the paper and reference implementation, owning the method, model, and evaluation.

Actions

• Designed the Trajectory-Signal Validity (TSV) test, which re-scores models on mid-spend customers against a naive threshold rule to expose hollow accuracy.
• Built TRACE, a transformer sequence encoder with self-supervised contrastive pre-training over short transaction windows and late fusion of transaction and engagement signals.
• Targeted the cold-start regime where aggregate RFM features are unstable, and published the work in IEEE format with a reference implementation.

Outcome

A reusable audit that separates genuine trajectory signal from re-identification, paired with a model built to pass it.