Harvard University
M.S. Data Science · Institute for Applied Computational Science
GPA 4.0/4.0. Coursework in ML, MLOps, Quantitative NLP, Multilevel Models, Big Data Systems. Research across RAG, AI safety, and interpretable health analytics.
Data Scientist · ML Researcher
Hi, I'm Joanna Wang.
Harvard Data Science candidate building retrieval-augmented and multi-agent ML systems for voice assistants, clinical research, and responsible automation.
-
Focus
Agentic ML, RAG, and AI safety for health + voice products
-
Latest
Amazon Alexa AI · Harvard & BCH CHIP Lab · ChronicQue
-
Impact
Cut Alexa log runtimes 40% and boosted RareMind coverage 13.6%
Currently
M.S. Data Science @ Harvard · May 2026
Studying at the Institute for Applied Computational Science with a 4.0 GPA—specializing in retrieval-augmented generation, model eval, and multi-agent orchestration.
- RAG + multi-agent orchestration (LangGraph/LangChain)
- LLM evaluation, causal inference, and AI safety
- Spark, Airflow, AWS (S3/EMR), BigQuery, Google Cloud
About
Shaping trustworthy AI, from bedside triage to household assistants.
My current work spans Alexa's Semantic Enrichment Pipeline, and ChronicQue's rare disease navigator. I love translating messy, high-volume signals into human-friendly insights with rigorous evaluation and strong product instincts.
Retrieval-Augmented Generation
Multi-agent Orchestration
Public Health AI
Bias & Safety Evaluation
Spark + BigQuery Analytics
Story-led Communication
Education
Academic foundations
Graduate and undergraduate training that shaped my approach to causal inference, large-scale systems, and responsible AI.
University of British Columbia
B.A. Economics & Statistics · GPA 4.0/4.0
Focused on statistical learning, causal inference, advanced econometrics, and time series analysis—fueling my product analytics toolkit.
Experience
Where I’ve shipped impact
From Alexa’s semantic cache to public-health intelligence, here are the teams and systems I’ve helped build recently.
Amazon · Alexa AI
Data Engineer, Alexa Semantic Enrichment Pipeline
Processed 18 TB/30-month utterance logs to support RAG for Alexa's semantic cache, cutting key joins from 3% mismatches to 0.08% and lifting tail-index retrieval to ~10M while reducing runtime 40% with salted Spark keys and AQE.
Harvard & BCH CHIP Lab · Prof. Maimuna Majumder
Machine Learning Engineer · Agentic Unified Review of Unstructured Media
Designed a multi-agent workflow that validates static web scraping with an LLM-backed extraction pipeline and modular roles (8 agents). Benchmarked against human-labeled datasets to deliver +100 pp discovery recall, +13.6% extra coverage from new events, 12% more fresh discoveries, and 29% fewer false positives after review, trading a small cost increase for broader, cleaner coverage.
Harvard IACS · ChroniCue
AI Researcher · Diagnostic Assistant for Rare & Common Diseases
Closed the 4–5 year rare-disease diagnosis gap with a hybrid engine where Phrank handles rare disorders alongside a medical LLM for common cases, all routed by a lightweight meta-learner.
Massachusetts Institute of Technology
Researcher · Context-Debias
Extended the Context-Debias framework to new attributes, adding L2 stability terms to preserve original semantic information; retained GLUE benchmark performance while reducing SEAT (Age/Disability) bias scores from 0.51 → 0.04.
British Columbia Lottery Corporation
Data Scientist, Product Analytics
Shipped a Looker Studio + BigQuery dashboard monitoring 300k+ user sessions, unlocking a 15% lift in engagement and training logistic models to flag session-level bounce risk, improving UX changes by 10%.
Selected work
Research, products, and publications
Recent systems I built or led—spanning RAG pipelines, health tech, and bias-aware NLP.
2024 — Present
Harvard University
Causal Inference of Wildfire Impact
Built a spatial panel linking NIFC fire perimeters to SafeGraph visits, indexing concentric buffers (0–2/2–5/5–10/10–25 km) around each alarm. Estimated a PPML difference-in-differences/event-study to recover the ATT: close-in visits fall ~40% immediately, while mid-band zones show a substitution bump before re-stabilizing six months later.
Ongoing
UBC Hydrology
Climate-Hydrology Multilevel Analysis
Built multilevel mixed-effects and Bayesian models linking watershed response to climate indicators, clustering basins to highlight shared risk patterns for water management.
Submitted
University of British Columbia
Network Meta-Analysis for Renewable Energy
Compared renewable energy technologies via Bayesian evidence synthesis and hierarchical modeling to rank investments and inform capital allocation.
Jun 2025 — Present
Harvard & BCH CHIP Lab
Agentic Unified Review Platform
Built a modular review stack for public-health surveillance with role-specific agents (Relevance Gate, Layout Parser, Fact Checker, Credibility Scorer, Arbiter) plus an LLM-validated extraction pipeline to transform static web scraping into structured alerts.
GitHub repo (going public soon)
Sep 2025 — Present
Harvard IACS
ChroniCue Diagnostic Engine
Hybrid matcher where Phrank scores rare diseases, a medical LLM handles common cases, and ClinPhen powers symptom extraction with negation handling, all orchestrated by a meta-learner for 20 ms responses at F1 0.80.
GitHub repo (going public soon)
Fall 2024
MIT · Context-Debias
Context-Debias Attribute Expansion
Extended the bias-mitigation framework with orthogonal loss terms, maintaining GLUE performance while shrinking SEAT Age/Disability scores 0.51 → 0.04.
Want to see prior geospatial + econometrics work? Browse the analytics archive for long-form notebooks and published analyses.
Hackathons
Building fast with thoughtful teams
Favorite sprints where collaboration, research, and storytelling came together.
2025 · Live AI
TurkEye
Multimodal deepfake moderation
ConvNeXt + DANN backbone with captioning and saliency to explain why an image is flagged. Designed to keep platforms safe while staying transparent.
View on GitHub
2025 · Harvard Rare Disease
RareMind
Patient-centric medical data portal
FastAPI + React platform that extracts HPO terms, triages symptoms via LLMs, and produces shareable diagnostic reports.
View on GitHub
2023 · ASA DataFest
BERT-powered insights
Legal Q&A prioritization
Blended BERT embeddings, TF-IDF, and SHAP to predict which pro bono legal questions get answered—achieving 87% accuracy.
Contact
Let's collaborate
I'm currently open to full time MLE/DS/DE opportunities on areas related to agentic system, responsible AI, or delightful data storytelling.