Full-Stack Data Scientist & Sr. Data Engineer — 20 years in biopharma manufacturing, 13 in data roles. I own the AI & analytics platform portfolio for a plasma fractionation site, anchored by an R/Shiny manufacturing analytics platform of ~84,000 production lines serving ~150 daily users on a 15-minute refresh cycle.
I build production GenAI that isn't allowed to guess — deterministic code computes every number; the language model is only allowed to narrate it.
- Eight years operating the plasma fractionation process as a manufacturing technician, then five analyzing it as an operations analyst — the floor-level process knowledge behind every model and platform I've built since.
- In a GxP environment, a chatbot that invents a batch yield isn't a bug — it's a compliance event. My GenAI architectures restrict the language model to interpretation while deterministic code computes all statistics; in production, that design also cut token consumption by ~70%.
- 30+ production solutions shipped, including 3 production RAG systems.
- Lean Six Sigma Green Belt · Databricks Champion · MuleSoft Champion
| Platform | What it is |
|---|---|
| AMIRA | Flagship R/Shiny manufacturing analytics platform — integrates batch historian, process historian, laboratory, and ERP data into one operational view for the site |
| Horizon | Quality-management dashboard over deviation and CAPA records — natural-language text-to-SQL chat, fiscal-year KPI scorecards with drill-downs, and (since v1.2.0) the hybrid RAG deviation search absorbed from the decommissioned DevSight app: retrieval over four years of quality records (2,213 records → 31,577 vector chunks) that cut investigation search from ~30 minutes to ~2 minutes |
| BioTrack | Manufacturing GenAI platform for microbial investigations — a deterministic nine-rule root-cause engine over manufacturing, laboratory, and ERP maintenance data (multiple-comparison correction, measured base rates), plus its own scoped RAG deviation search built on the same retrieval engine as Horizon's |
| TOMO | Multimodal conversational AI assistant embedded in the site's analytics platforms — streaming responses, vision, and voice |
These run in a regulated environment on my employer's private GitHub Enterprise, so the code can't be public — happy to talk architecture.
Impact highlights: a multivariate deep-dive into an automated filtration process enabled a ~25% cycle-time reduction worth over $2.0M per year · text analytics and NLP cut ERP work-order search effort by 75%.
| Project | What it is | Link |
|---|---|---|
| Shiny Cost Estimator | Interactive COCOMO II cost estimator for R Shiny and data science projects — analyze a repo or enter line counts, get cost/schedule/team estimates and a branded PDF report (v2.0.0, MIT — 42% cost cliffs eliminated) | Repo |
| Aether Wardrobe | AI wardrobe assistant — a deterministic TypeScript engine generates outfit combinations; the LLM only scores and explains them (Next.js, React 19, Supabase) | Repo · Live app |
| {RIOT} | R/Shiny Image OCR Technology — upload an image, interactively select regions, and extract editable text in real time via the Tesseract engine | Repo |
| {roldanpack} | Personal R package of helpers for daily analytics and visualization work | Repo |
| TubeScout | Python/Streamlit app that surfaces high-signal YouTube videos independent of the algorithm by analyzing view-to-subscriber ratios | Repo |
| MyRepoBot | Python CLI that automates GitHub repo and local project setup, with templates for R/Shiny projects | Repo |
| Portfolio site | Next.js/React portfolio — one of 8 production TypeScript apps | Site |
More experiments and older projects live on my repositories page.
Expert — 
Advanced — 
Applied —
Platform — 
