PhD, Industrial & Systems Engineering
University of Washington
Explore my question universe
Click a question to bring it to the centre — then click the centred one to open its project.
Retrieval quality starts with content quality and understanding.
The goal of AI is to create simplicity and automation — not to expose more AI to the user.
Complex AI systems need evaluation beyond either humans or LLM judges alone.
Deep models over audio, protein sequences and user graphs — turning raw signals into predictions.
Journey
2019
Wyze Labs
Research Scientist II, Audio AI
2021
Warner Music Group
Data Scientist, Entity Matching & Music AI
2022
Microsoft Research
Applied Scientist II, NLP & Large Language Models
2023
AstrumU
Lead Applied Data Scientist, Skill-Based Education & Hiring
Now
ProRata.ai
Principal AI Engineering Manager
// Experience
Experience
Teams and problems I’ve worked on — from internships through research to production AI.
View full résuméEducation
MSc, Statistics
University of Washington
BS, Statistics — with Distinction
University of Iowa
// Projects
Hard problems, and what I built to solve them
Grouped by the kind of problem. Pick a question to see the system behind it.
1 / Pick a problem
Multimodal Retrieval & RAG
Agentic Automation
System Evaluation
Other Intelligences
2 / How I solved it
-
ProRata.ai
Knowledge Graph / OntoKG
There are thousands of Donald Trumps. Are you referring to the president?
Design the taxonomy first, then extraction pipelines that scale to all of them.
ProRata.ai
Knowledge Graph / OntoKG
Design the taxonomy first, then extraction pipelines that scale to all of them.
Knowledge Graph
- Designed entity and relation taxonomies spanning people, organizations, places, events, creative works, products, science, and knowledge domains.
- Built a large-scale Wikidata-derived entity index, reducing 100M+ raw entities to ~34M high-quality canonical entities through filtering, normalization, and cleanup.
- Constructed an entity property graph capturing both intrinsic relationships and topical relationships to support retrieval, reasoning, entity discovery, and contextual understanding.
- Developed large-scale entity extraction, linking, and expansion pipelines over publisher content, grounding extracted entities against the canonical entity index and expanding entity coverage to 100M+ entities and entity mentions.
- Productionized entity services for Gist Answers, search, and GEO applications, enabling entity-aware retrieval, query understanding, contextual reasoning, and content discovery.
Basis of the OntoKG paper submitted to ISWC 2026.
-
ProRata.ai
Discovery Questions
What if I already knew what you’d ask next?
Generate the questions from the content, and retrieve them at scale.
ProRata.ai
Discovery Questions
Generate the questions from the content, and retrieve them at scale.
Discovery Questions / Gist Widget
- Built LLM-based systems to automatically generate contextual discovery questions from publisher content for Gist search and discovery experiences.
- Deployed the Gist Widget across 30+ publishers, reaching 20M+ monthly widget views, approximately 0.5% engagement, and generating 65K+ monthly Gist queries.
- Developed an iterative quality-improvement framework combining LLM-based evaluation, human-in-the-loop validation, and online CTR feedback to improve generated-question relevance and engagement.
- Established evaluation datasets and quality metrics to identify generation failure modes and continuously improve question-generation pipelines.
20M+ monthly widget views and 65K+ monthly Gist queries across 30+ publishers.
-
ProRata.ai
Multimodal & Music Attribution
How should Suno pay UMG and music artists?
Understanding and matching across text, tables, images, audio and video.
- Music Business Worldwide — Universal Music × ProRata
- Interactive demo coming soon.
ProRata.ai
Multimodal & Music Attribution
Understanding and matching across text, tables, images, audio and video.
Music Attribution / Generative Music
- Developed a retrieval-based music attribution algorithm for identifying reference music associated with generative-AI outputs using music-industry reference catalogs.
- Built retrieval and matching pipelines for large-scale music comparison and attribution.
- Improved retrieval quality while optimizing system architecture to reduce inference latency and computational cost.
- Collaborated with music-industry stakeholders on reference-data and attribution requirements.
Multimodal Attribution
- Led multimodal retrieval and attribution initiatives spanning text, images, audio, and video.
- Developed systems for multimodal content understanding, retrieval, matching, and source/provenance attribution.
- Integrated multimodal intelligence into publisher-facing and search-oriented AI applications.
Includes a submitted patent on multimodal attribution.
-
ProRata.ai
PACE
Your RAG system just quoted the cookie banner. Now what?
An agent learns each publisher’s layout once, then extracts with zero LLM cost.
ProRata.ai
PACE — Publisher-Adaptive Content Extractor
An agent learns each publisher’s layout once, then extracts with zero LLM cost.
Goal
- Build a high-quality webpage extraction system for RAG that removes page noise while preserving multimodal content such as text, tables, and images.
Approach
- Developed an agentic, publisher-adaptive extraction system with a two-stage training and inference architecture.
- During training, PACE uses a limited number of LLM calls plus human-in-the-loop feedback to learn publisher-specific extraction patterns.
- The agent decides whether to reuse an existing extractor or create a new rule-based extractor for a publisher.
- During inference, extraction runs with zero LLM cost and low latency.
- Supports multiple output formats, including plain text and Markdown, while preserving structured and multimodal content.
Impact
- Achieved state-of-the-art performance on webpage content extraction and noise removal.
- Deployed across 1,500+ publishers.
- Produces high-quality, RAG-ready content at scale with effectively zero LLM inference cost.
-
ProRata.ai
EPOCH
Are you still tuning prompts by hand?
One evaluation-driven loop over prompts, rules, fine-tuning and code.
- GitHub repository — EPOCH
- arXiv 2603.09049 — EPOCH: An Agentic Protocol for Multi-Round System Optimization
ProRata.ai
EPOCH
One evaluation-driven loop over prompts, rules, fine-tuning and code.
EPOCH — Agentic Prompt & System Optimization
- Architected and developed EPOCH, an internal agentic optimization platform that integrates task-specific evaluation with automated experimentation.
- Built workflows for automatically establishing baselines, modifying prompts, rules, configurations, and code, executing evaluations, and retaining changes that improve target metrics.
- Applied EPOCH to production tasks including query understanding and discovery-question generation, reducing manual prompt iteration and enabling evaluation-driven system optimization.
- Contributed to the associated research publication on systematic agentic optimization.
Basis of a research publication submitted to COLM 2026.
-
ProRata.ai
Gist Answer + LLM Eval
Who checks the AI that’s checking the AI?
Score it against human-feedback benchmarks, not just another model.
ProRata.ai
Gist Answer + LLM Eval
Score it against human-feedback benchmarks, not just another model.
Gist Answer & LLM Evaluation
- Designed LLM-as-a-Judge and human-in-the-loop evaluation frameworks for complex generative AI systems.
- Developed binary preference and multiple-choice annotation tasks to validate automated evaluation metrics against human judgments.
- Used human agreement, disagreement, and confidence signals to improve the robustness, calibration, and reliability of LLM evaluators.
- Applied the evaluation framework across tasks including text denoising, chunking, question generation, and answer generation.
Evaluation pipeline grounds answer quality in human feedback benchmarks.
-
AstrumU
SkillLLM & Skill Taxonomy
Your résumé says PyTorch. The job asks for deep learning. Match?
Fine-tuned skill extraction on top of a comprehensive skill taxonomy.
- arXiv 2410.12052 — Skill-LLM: Repurposing General-Purpose LLMs for Skill Extraction
- Press release — ETS, JFFLabs & AstrumU partnership
- AstrumU — Enterprise Organizations overview (PDF)
AstrumU
SkillLLM & Skill Taxonomy
Fine-tuned skill extraction on top of a comprehensive skill taxonomy.
SkillLLM
- Fine-tuned Large Language Models for contextual skill extraction from job descriptions and professional profiles.
- Productionized large-scale extraction pipelines processing millions of job descriptions and profiles, converting unstructured text into normalized skill representations.
- Improved skill identification and extraction quality through domain-specific training, evaluation, and model iteration.
Skill Taxonomy
- Architected a large-scale skill taxonomy and knowledge graph spanning technical and soft skills.
- Built automated pipelines to normalize, deduplicate, link, and expand extracted skills into canonical taxonomy concepts.
- Designed a continuous taxonomy-expansion process using newly extracted skills and semantic relationships from workforce data.
Skill Matching & Recommendation
- Developed retrieval- and embedding-based systems for skill normalization, disambiguation, and semantic matching.
- Built skill-matching capabilities supporting job recommendation, skill-gap analysis, career intelligence, and workforce analytics.
- Integrated SkillLLM, the skill taxonomy, and semantic retrieval into end-to-end production recommendation systems.
AI Team & Platform
- Scaled a multidisciplinary AI team from 2 to 7+ members and led end-to-end AI development spanning data annotation, model training, evaluation, deployment, and production inference.
- Owned technical architecture and delivery across multiple AI-powered career intelligence and recommendation products.
-
Microsoft Research
Science Engine / Antigen LM
A language model that reads proteins, not prose.
A protein language model over 2B+ sequences, scoring how natural it looks.
Microsoft Research
Science Engine / Antigen LM
A protein language model over 2B+ sequences, scoring how natural it looks.
Science Engine / Antigen Language Model
- Trained BERT-based antigen/protein language models on 2B+ multi-species sequences using distributed multi-GPU training.
- Improved model perplexity from 1.7 to 1.4, outperforming the state-of-the-art AntiBERTy model.
- Developed naturalness-scoring methods for identifying atypical antigen sequences and supporting downstream biological discovery.
Scientific Entity Extraction
- Developed transformer-based entity extraction systems for scientific papers and patents.
- Benchmarked fine-tuned transformer approaches against GPT-2 and GPT-3 few-shot learning across multiple scientific NLP tasks.
GPT Document Intelligence
- Built GPT-based pipelines for summarization, entity-relation extraction, question answering, and document generation.
- Fine-tuned GPT models on European Medicines Agency datasets for domain-specific document intelligence applications.
-
Warner Music Group
Music Recognition & Music AI
Is there a new trend of music?
String matching plus learned embeddings — and genres never seen before.
Warner Music Group
Music Recognition & Music AI
String matching plus learned embeddings — and genres never seen before.
Music Copyright & Entity Matching
- Developed large-scale music copyright identification and metadata reconciliation systems.
- Combined deterministic and fuzzy string-matching techniques with deep-learning entity embeddings to resolve noisy artist, recording, and catalog metadata.
Open-Set Music Recognition
- Developed deep-learning models for music genre classification and trend analysis.
- Researched open-set recognition methods capable of identifying emerging or previously unseen music genres outside a predefined taxonomy.
Lyric Recognition
- Collaborated with Warner Records on deep-learning systems for lyric recognition and restoration.
- Developed models for recovering and generating lyrics from historical music recordings.
-
Wyze Labs
Audio Event AI
Is your baby crying when you are not in the room?
Deep audio detection in the cloud, plus efficient C++ on the device.
Wyze Labs
Audio Event AI
Deep audio detection in the cloud, plus efficient C++ on the device.
Audio Event Dataset & Annotation
- Led the organization and management of annotation workflows for large-scale sound-event datasets sourced from public audio and customer feedback.
- Designed annotation guidelines, tutorials, quality metrics, and evaluation processes to ensure consistent high-quality training data.
Cloud Audio Event Recognition
- Researched and developed deep-learning audio classification models for complex sound-event recognition.
- Collaborated with engineering teams to deploy the models as scalable cloud services for Wyze cameras.
- Enabled detection of important events including baby cries, dog barks, and gunshots.
Edge Audio Event Detection
- Partnered with edge-device engineers to develop an efficient C++ on-device audio event detection algorithm.
- Designed filtering logic that removed 95%+ of irrelevant/noise audio events during testing, reducing unnecessary downstream cloud processing.
-
Alibaba Group
GCN Recommendation
How does Alibaba know what you’ll like before you do?
Treat it as a graph and let multi-view convolutions fill the gaps.
Alibaba Group
GCN Recommendation
Treat it as a graph and let multi-view convolutions fill the gaps.
MV-GCN Recommendation / Link Prediction
- Developed MV-GCN, a Multi-View Graph Convolutional Network, for link prediction and recommendation based on graph matrix completion.
- Modeled interaction information and content information as separate graph views rather than directly concatenating heterogeneous features.
- Introduced cross-view embedding consistency constraints to improve learned graph representations and prediction accuracy.
- Evaluated the approach across homogeneous and heterogeneous recommendation datasets, outperforming contemporary graph-based matrix-completion baselines.
// Blog
Blog
Notes on AI, engineering, and things I’m learning.
// Recognition
Recognition
Awards, patents, and publications that mark milestones in my work.
Award
Best Refereed Paper, Finalist
2019
For “Identifying the Influential Inputs for Network Output Variance Using Sparse Polynomial Chaos Expansion”
Granted
Cross-Document Intelligent Authoring And Processing, Including Format For Semantically-Annotated Documents
US20210081601A1
Submitted
Multimodal Attribution
Submitted
Systems and Methods for Enhancing Content Navigation and Discovery
Draft
Agentic Content Understanding for RAG and GEO
Rendered from data/publications.js — the complete record is also on
Google Scholar.
// Contact
Get in touch
Have a question, an opportunity, or just want to say hi? I’d love to hear from you.