💼 Analytics Engineer | Decision Systems | BI Developor
🌏 Based in Seattle, WA
I'm a data-driven problem solver with a background in Informatics: Data Science and a Master's in Analytics. I thrive on translating complex business questions into actionable data solutions, leveraging my expertise in causal inference and decision science to drive impactful outcomes. With a passion for building scalable analytics pipelines and fostering a metric-driven mindset, I am dedicated to empowering teams to make informed decisions through data. Based in Seattle, I am always eager to connect with fellow data enthusiasts and professionals in the field.
🤝 Connect with me : 📫 LinkedIn | ✉️ nsnatthapat@gmail.com
Professional Case Study
Implemented an enterprise performance management (EPM) solution using Board to support financial planning and reporting across multiple departments.
Context
The organization required a centralized platform for budgeting, forecasting, and operational reporting to replace fragmented spreadsheet-based workflows.
Contribution
- Designed multidimensional data models supporting finance and operational reporting
- Built dashboards and planning workflows for budgeting and performance tracking
- Integrated SQL-based data sources and flat files into Board cubes
- Supported stakeholders in Finance and PMO teams through reporting and analysis needs
Tools
Board EPM • SQL • Excel • Data Modeling • OLAP Cubes
Outcome Enabled centralized planning and reporting workflows, reducing reliance on manual spreadsheet consolidation and improving visibility into operational performance.
Professional Case Study
Developed operational reporting solutions using Oracle Transactional Business Intelligence (OTBI) to support business monitoring and decision-making.
Context
Business teams required improved access to operational metrics from Oracle systems without relying on manual reporting or IT-managed data extracts.
Contribution
- Designed and developed OTBI reports for operational and financial monitoring
- Built custom analyses combining multiple subject areas within Oracle data models
- Collaborated with stakeholders to translate business questions into structured reporting logic
- Documented report logic and trained users on interpreting key metrics
Tools
Oracle OTBI • SQL • Oracle ERP Data Models
Outcome Enabled business teams to access self-service reporting directly within the Oracle ecosystem, reducing turnaround time for operational insights.
Causal Inference, Counterfactual, Channel Attribution and ROI Analysis
| Project Link | Project Description |
|---|---|
| Multi-Channel Online Ad Effectiveness Analysis - The Criterion Channel | Analyzed a 1.2M-user randomized experiment across five ad channels (Facebook, Twitter, Instagram, YouTube, Letterboxd) to measure incremental subscription lift and channel-level ROI. Found ~348% relative lift overall, with Twitter as the only statistically reliable positive-ROI channel and Instagram as a promising but underpowered opportunity. |
| Online Ad Campaign Effectiveness Analysis - Rocket Fuel | Evaluated a 96/4 treatment-control ad experiment using unadjusted lift, covariate-adjusted OLS, IPW, and Double Machine Learning. Found ~42% relative conversion lift and a treatment-only ROI of 37.7%, with actionable segmentation by impression frequency, day of week, and hour of day. |
| Geo-Randomized Paid Search Effectiveness Analysis - eBay | Used a DMA-level ad suspension experiment and two-way fixed effects Difference-in-Differences to contrast naïve attribution (+272% ROI) against the true causal return (−60% ROI), quantifying how organic substitution causes standard attribution to overstate paid search value by over 330 percentage points. |
| Multi-Channel Campaign Attribution & ROI Analysis - Patagonia | Compared last-touch, first-touch, and linear attribution against intent-to-treat causal estimation across five channels (email, YouTube, Instagram, Google Search, contextual display) on 2M customer records. Found naive models overstate marketing impact ~4x, with contextual display flipping from the top-ranked channel (naive) to negative ROI (incremental). |
| Discount Coupon Email Effectiveness, Targetting Method, and ROI Analysis - Artea | Ran A/B test analysis on a 5,000-customer coupon experiment and built a fully interacted heterogeneous treatment effects model to identify high-lift segments (cart holders, Instagram-acquired users). Surfaced an algorithmic fairness issue: the targeting model significantly under-routes coupons to minority customers via acquisition channel as a proxy variable. |
| Customer Response Ranking and Marginal ROI Analysis - Electronic Arts | Trained CART and XGBoost models to rank mobile ad impressions by predicted CTR, finding that historical user CTR dominates in-session ad variety as a targeting signal. Showed targeted buying of top-ranked impressions nearly doubles ROI (351% → 682%) and derived the optimal mixed-channel budget allocation using a marginal ROI crossover framework. |
| Cross-Brand Customer ID consolidation measurement on Marketing ROI: A Customer Data Platform Analysis - Amperity | Simulated fragmented vs. resolved customer identity across a multi-brand retailer and quantified downstream distortion: 51% customer count overstatement, high-value customer share halved, and a 33-point ROI gap between realized performance and what the fragmented system self-reported — illustrating how poor identity quality creates both targeting errors and false measurement confidence. |
| Display Advertising Effectiveness A/B Test - UberEats | Compared ITT and ATT estimators across 576K users to measure display ad impact, finding revenue lift concentrated in non-American markets and dormant users, while active users showed no significant response. Demonstrated that a naive within-treatment comparison inflates the estimated lift by ~30x due to selection bias in ad exposure. |
| Geo-Randomized Ad Removal A/B Test - Starbucks | Applied two-way fixed effects DiD on a 200-market panel to estimate the causal impact of Google Maps local search ads, finding a statistically significant lift of ~9.6 units per hour per store (~7.4% relative lift). Validated parallel trends via event study and pre-period regression before interpreting the estimate as causal. |
| Project Link | Project Description |
|---|---|
| Retail Market Research: Consumer Segmentation with PCA | PApplied PCA to a retail attitude survey, reducing five correlated preference items into two interpretable dimensions — service orientation and price sensitivity — that together explain 90.5% of variance. Mapped respondents into four actionable customer segments to inform discount vs. full-service retail positioning decisions. |
| Project Link | Project Description |
|---|---|
| Market Research: Synthetic Customer Ratings via LLM - Wine Reviews | Benchmarked zero-shot, few-shot, and chain-of-thought prompting strategies for LLM-based rating of wine reviews against human critic scores. Zero-shot (MAE 1.71) outperformed all engineered variants, while inter-run variance analysis (7-point spread across 20 identical calls) revealed non-determinism as the primary production risk for LLM annotation pipelines. |
| LLM Fine-Tuning for Customer Review Triage - Amazon Reviews | Fine-tuned GPT-4.1-mini on 67 labeled magazine subscription reviews to produce structured JSON triage outputs (complaint tag, escalation flag, suggested action). Fine-tuning improved complaint tag accuracy from 27.3% to 63.6% and JSON validity to 100%, demonstrating that task-specific fine-tuning outperforms prompt engineering for schema-constrained classification. |
| RAG System for Academic Program Office - UW MSBA | Built and evaluated a retrieval-augmented generation system over 21 MSBA program documents (syllabi, policies, handbooks) for student support use cases. RAG scored 40/40 vs. 18/40 for the plain LLM baseline across factual accuracy, specificity, relevance, and hallucination control — with the largest gains on high-stakes policy and deadline queries. |