What I build
End‑to‑end data + AI stacks for real‑world monitoring: ingestion, normalization, modeling, geo/time signals, and delivery — with provenance and evaluation wired in.
Personal page • applied AI for social good
Building data + AI systems for social good.
I build monitoring‑grade pipelines that turn messy signals into traceable evidence, decision‑ready products, and ultimately, actionable insights. The work is systems‑first: research design, deep learning + information retrieval (RAG/KB patterns, AI‑assisted), algorithm design, geo + time analytics, and evaluation harnesses that stay attached in production.
Blueprint at a glance
A reusable workflow I apply across social‑impact domains.
monitoring‑grade
R&D: RQ → op-sigs → baselines → fail modes
arch: conn → parse → schema/patch → lineage
K/R: embed → vec+KB → hybrid RAG
mdl: multi-head → cal+unc → decision policy
LLM: synth rules | assist labels | doc→schema | err triage
ship: agg → dash/maps → alerts/briefs | API
SRE: drift | health | audit | ver | rollback
Evidence‑linked
Comparability
Continuous eval
Audit trails
Human review
Design systems that can pivot across countries and crises—because today’s politics is volatile, and public-interest tools need to be portable and accountable.
I care about who gets protected by the system: contexts that are under-resourced, over-scrutinized, or politically targeted. That means stress-testing failure modes, tracking drift, and keeping outputs explainable—especially when power wants ambiguity.
About
When politics turns coercive, we don’t have to accept it.
How I work
Calm surfaces over hard problems: schema stability, robust pipelines, algorithm design, and continuous error analysis so outputs stay usable under noise and drift.
Collaboration
If you’re building mission‑driven products (research labs, NGOs, policy, or industry), I’m happy to talk. Low‑friction contact here.
Deployments
A few deployed configurations that show how the same stack is parameterized across problem settings.
Deployed configurations
Civic space monitoring & forecasts
Monitoring
Event detection + interpretable early‑warning signals for shifts in protest, restriction, media pressure, and advocacy activity — designed for frequent refreshes and evidence traceability.
Foreign influence & coercive leverage tracking
Influence
Tracks influence patterns across channels (diplomatic / economic / information / cyber) with consistent task definitions, careful normalization, and transparent aggregation.
Climate‑driven disruption & response signals
Climate
Detects environmental shocks and social responses using multilingual event modeling + geo grounding, with downstream time‑series signals for monitoring and analysis.
Subnational disruption monitoring (ADM1)
Subnational
Map‑first monitoring at subnational resolution, wired to two‑stage geo reconciliation (country → ADM1), standardized counts, and surge detection for rapid scanning and drill‑down.
These are representative, not exhaustive — the emphasis is on reusable infrastructure that transfers to new domains and stakeholders.
Selected papers & technical notes
Previews only (first 2 pages) — to avoid circulating full drafts. Public versions are linked when available.
Tracking Civic Space in Developing Countries with a High‑Quality Corpus of Domestic Media and Transformer Models
Foreign Influence by Authoritarian Governments: Introducing New Data and Evidence
#DataForUkraine: Adapting Social Science Tools for Crisis Response
Benchmarking Causal Inference Methods for ATE Estimation
Modular Gated Attention: Adaptive Architecture for Flexible Sequence Modeling
If you’re exploring collaborations, I can share additional technical notes (evaluation harnesses, sampling QA playbooks, schema/patch patterns).
Impact themes
Typical problem spaces these systems support.
Civic space & democracy
Information integrity
Foreign influence
Climate risk & adaptation
Humanitarian response
Auditability
Links
Low‑friction ways to reach me or explore artifacts.
Contact
Best for collaboration notes, dataset/tool questions, and applied AI opportunities.
Code & models
Pipelines, experiments, and artifacts that support monitoring‑grade AI systems.
Lab
Where a lot of the public‑interest systems work lives.
Small things
Useful work still benefits from a little life.
About me
Taiwanese by origin, culturally curious by default. I travel through food and the stories people tell through what they cook. Off the clock: dancing, handcrafts, and graffiti-inspired lettering—creative outlets that keep me energized.
Quick nav
Technique palette
Multi-task / multi-head modeling
Transformers + LLM integration
Retrieval (hybrid search, RAG)
Feature / embedding engineering
Evaluation harness + error analysis
Serving + latency-aware pipelines
Observability + drift monitoring
Data contracts + schema/patch layers
