Faye Xia

I’m Fei, a CMU Chemical Engineering student and AI startup co-founder. I’m drawn to unmet user needs and the data that reveals them, and enjoy turning open-ended questions into practical technical solutions that make everyday experiences better.

Experience

Procter & Gamble · Cincinnati, OH · May–August 2026

Digital Innovation InternOral Care R&D

Connected stakeholder needs with engineering implementation for an R&D knowledge chatbot and automated data pipeline spanning 5,000+ research documents.

More detail
  • Surfaced stakeholder needs for an R&D knowledge chatbot through interviews and stakeholder sessions; clarified expectations with my manager and coach and translated requirements into technical questions for engineering discussions.
  • Coordinated with a Europe-based engineer and scheduled user testing so feedback could inform subsequent technical iterations.
  • Evaluated an existing RAG pipeline for domain-specific retrieval, informing the choice of Databricks-native retrieval and text-to-SQL tools.
  • Built an automated pipeline (ClaimSync) to scrape, parse, and extract structured codes from 5,000+ research documents; loaded results to Databricks to power downstream search and AI retrieval.
ClaimSync
Automated collection & parsing of 5,000+ R&D claims documents into a structured database
① Structure the data
→
Claims Center
Power BI dashboard over joined POA + POC data — 13 filter dimensions
② Make it searchable
→
Claims AI
GPT-4.1 orchestrator with dual RAG + Text2SQL retrieval via a DashApp frontend
③ Make it conversational

Three successive systems built over Summer 2026 at P&G Oral Care R&D · Cincinnati, OH

① Download Review Docs
Automated HTML scraping of Claims Manager website; downloads POC review documents
→
② Download Attachments
Parser opens each document and extracts embedded POC .docx attachments
→
③ Find FOP Codes
Extracts & characterizes 8-digit codes — FOP, PMP, POA, FPC — using context rules

All three steps run end-to-end via CLI: gcas-interactive

POA Data
Databricks
+
POC Data
via ClaimSync
+
SharePoint URLs
attachment links
↓  KNIME workflow — join on FOPs  ↓
Final Data Table
Enriched columns: POA# · FPC life stage · Segments · Regions · Brand / sub-brand · Market claims
↓
Power BI Dashboard
Oral Care Claims Center · 13 filter dimensions
Databricks
stored for downstream Claims AI retrieval

KNIME joins three sources on 8-digit FOP codes; Power BI enables cross-filter search by FOP, FPC, Brand, Segment, Region, and more

User Question
↓
GPT-4.1 Orchestrator
  • Safety screening
  • Question classification — 4 types: Structured Retrieval, Semantic Discovery, Analytical Comparison, Support Generation
  • Parallel retrieval across both backends
  • Answer synthesis
↙             ↘
Vector Search (RAG)
Databricks · semantic discovery of claims documents
Genie / Text2SQL
Databricks · structured retrieval from claims database
↓ synthesize ↓
Response
Grounded in both semantic matches and structured data rows

Frontend: DashApp · Deployed on Databricks · GPT-4.1 routes each query to the right retrieval backend

stakeholder coordinationuser testingRAGDatabricksSQL
Mafold · July 2026–present

Co-founderAI agent product development

I work on turning user needs into product priorities, technical requirements, and clearer ways for a remote team to collaborate.

More detail
  • Identified a gap in sharing personal-agent work across teams; proposed agent-to-agent collaboration and broke the concept into technical tasks and privacy requirements with a backend engineer.
  • Co-defined a proposed access-control workflow with owner-managed allowlists and owner-consent escalation for sensitive requests.
  • Helped establish weekly team syncs and shared issue reporting across a 12-hour time difference, with agent-first troubleshooting and backend escalation.
  • Identified 12 product gaps across privacy, orchestration, onboarding, and model routing; organized triage using 3 issue categories and 5 priority levels.
productmulti-agentrequirementsremote coordination
Carnegie Mellon University · December 2025–present

Kitchin Lab ResearcherViscosity prediction & laboratory automation

I investigate whether ML models answer the scientific question we actually care about, and connect experimental automation with predictive modeling.

More detail
  • Viscosity benchmark audit: Audited 39,077 measurements; standardized 1,733 molecular representations into 1,367 unique molecules and found training overlap in 27 of 65 test molecules.
  • Showed structure-based models outperform structure-free learned embeddings for viscosity prediction (RMSE 0.060 vs 0.105 on held-out mixtures; 0.127 vs 0.314 on unseen molecules) using custom train/test splits, a physics mixing-rule baseline, and multi-seed checks; separately, RDKit molecular descriptors cut error by 56% over structure-free embeddings (RMSE 0.110 → 0.048).
  • Designed and executed 8 MixEmbed robustness tests, identifying 7 defects involving input validation, missing-data handling, and dependency compatibility. Findings communicated to the library maintainer resulted in fixes for all 7 issues.
  • Opentrons Flex: Automated 96 dye mixtures at 4 wavelengths. Trained regression models on 76 wells and evaluated 20 held-out wells (R² 0.9912–0.9989; MAE 0.0106–0.0296 OD).
viscosityML evaluationregressionOpentrons FlexPython
Carnegie Mellon University Robinson Lab Sep 2024 – May 2026

Research Assistant Protein purification & yeast transformation

Designed and led troubleshooting of a yeast and mammalian protein purification protocol across 7 controlled factors, raising protein yield by 200%.

More detail
  • Validated results with gel electrophoresis and Western blot.
  • Ran collaborative yeast transformation experiments targeting GPCRs using nanodisc technology; funded by a CMU ChESS Research Fellowship and presented at the Berg Scholars Symposium.
wet labWestern blotyeastGPCRs
Guangzhou Doublle Bioproducts Co. Ltd Guangzhou, CN Jun – Sep 2023

Research Intern Biomaterials for wound healing

Engineered hydrogels incorporating stem cell-derived exosomes to improve wound-healing efficiency; characterized structure and protein content with TEM and Western blotting.

More detail
  • Conducted in vivo wound-healing studies in mice and analyzed longitudinal biological data.
  • Co-inventor on the 2024 patent “Stem cell exosome-based hair restorer and its application.”
patenthydrogelsTEMin vivo

Projects

L’Oréal Brandstorm 2026 · November 2025–February 2026 · Initial-stage participation

ÉCHOlogneMarket research & product concept development

Helped shape a Prada-aligned fragrance concept: a refillable dual-chamber system paired with an NFC-connected app for scent discovery, visualizations, and music.

More detail
  • Conducted market and competitor research to inform brand selection and identify opportunities in personalization, brand identity, and differentiation from lower-cost alternatives.
  • Contributed to feature definition, pricing, and a phased rollout roadmap for the proposed product.
  • The deck presents a competition concept and proposed business plan. Financial forecasts and sustainability figures are targets or estimates, not achieved operating results.
productmarket researchpricingroadmap
Carnegie Mellon Data Visualization (06-643), graduate course Jan – Feb 2026

Early Yeast Culture Outcome Prediction

Can the first four hours of a yeast growth curve predict final GFP productivity? Engineered kinetic features from 75 samples across three carbon sources (YPD, YPG, YPDG) and compared symbolic regression, Random Forest, and XGBoost.

More detail
  • Features: initial biomass, maximum growth rate (μmax), lag time, early biomass accumulation (AUC4h), and curvature; raw OD smoothed with a Savitzky–Golay filter after blank validation and abnormal-well removal.
  • Symbolic regression on log-transformed productivity reached R² = 0.9876 and MSE = 0.1023 on 23 held-out samples; t-SNE showed media type structures early growth phenotypes.
symbolic regressionXGBoostfeature engineeringPythonyeast
Carnegie Mellon Creating Scientific Software (06-643) Spring 2026

Mana · AI study scheduler for Canvas

A Python CLI that pulls Canvas assignments and Google Calendar events, uses Gemini to classify deadlines and build a daily study plan around your habits, and sends email and SMS reminders through the Gmail API. Iterated through five versions.

More detail
  • Natural-language schedule edits (mana schedule "add gym at 3pm"), submission-status indicators from the Canvas Submissions API, fuzzy deduplication of calendar and Canvas items, and a daily cron job.
  • Interactive setup command with live credential validation; 100+ pytest tests and GitHub Actions CI.
Gemini APIGoogle OAuthPythonpytestCI
Carnegie Mellon Fluid Mechanics term project Spring 2026

Inclined Venturi Meter for Fertilizer Irrigation

Used SolidWorks and COMSOL in a four-person class project to design and analyze an irrigation-inspired Venturi meter operating at a 45° incline; compared simulations and theory with physical prototype measurements.

More detail
  • Geometry: 160 mm length, 18 mm inlet/outlet, 6 mm throat, 135° elbow, and angled manometer bases so the tubes stay vertical on the incline; the team used a resin-printed prototype with a target budget below $25.
  • Compared Bernoulli/continuity theory, COMSOL simulation (~0.6 m/s centerline throat velocity), and bench measurements (1649 Pa inlet → 472 Pa throat); the ~785 Pa hydrostatic term proved small relative to velocity-driven pressure change.
SolidWorksCOMSOL3D printingfluid dynamics
Carnegie Mellon · May 2025

Monument Valley Game Replica15-112 term project · Group leader

Led a student team to develop a playable Python game selected for the Top 20 Projects Gallery out of 150 projects.

More detail
  • Implemented 3D-to-2D projection, matrix transformations, interactive rotation, and screen transitions.
Pythongraphicsteam leadership

Skills

Programming & data

PythonNumPypandasscikit-learnseabornSQLC/C++MATLABGitDatabricks

Machine learning

LLM orchestrationRAGtext-to-SQLregressionsymbolic regressionXGBoostfeature engineeringcross-validation

Design & simulation

SolidWorksCOMSOL Multiphysics3D printingOpentrons Flex

Wet lab

protein purificationWestern blotgel electrophoresisyeast transformationTEMin vivo studies

Honors

★
Dean's ListFall 2024 and Spring 2026
◆
CMU ChESS Research Fellowship$4,500 (2025); presented at the Berg Scholars Symposium
✦
Patent co-inventor“Stem cell exosome-based hair restorer and its application” (2024)
▲
Top 20 Projects Gallery15-112, Monument Valley replica (2025)

Education

Pittsburgh, PA Expected May 2028

Carnegie Mellon University B.S. Chemical Engineering · Minors in Biomedical Engineering and Machine Learning

Relevant coursework: Artificial Intelligence and Machine Learning I; Computational Optimization and Machine Learning for Chemical Engineering; Probability Theory for Computer Scientists; Principles of Imperative Computation.

Contact

Your search — — did not match anything on this page.
Try “LLM”, “yeast”, “SolidWorks”, or clear the search.