Route with intent
Use the smallest capable path and make delegation visible.
Portfolio / 2026
The Applied AI Systems Edition
A five-chapter field guide to how I turn models, data, and product constraints into dependable systems for real decisions.
Jinghang Li, Ph.D.
Chapter ISystems that converse, reason, and act
Agentic AI
LLM systems built around routing, grounding, evaluation, and human judgment.
See the systemFind the best next action for this urgent service request.
Checking context and policy
Grounded response
I found three qualified options, verified availability, and ranked the tradeoffs.
✓ Sources checked ✓ Policy passed
Confidence
0.94
Human review available
01 / Production AI
Fluency is easy. Dependable behavior is the design problem.
I connect orchestration, semantic knowledge, prompt systems, and evaluation loops so an agent can make progress without hiding uncertainty from the people using it.
700K+monthly AI interactions in the system context
4layers: route, retrieve, reason, review
1goal: a useful next action
Ground every claim
Bring policy, knowledge, and operational context into the response.
Evaluate behavior
Test usefulness, risk, and consistency—not only model fluency.
Keep people in control
Expose confidence, reversibility, and a clear review path.
Chapter IIFrom a large network to the next best match
Recommendations
Scoring systems that combine fit, constraints, availability, and real operating context.
Follow the signal
Incoming request
Service need
Skill · location · timing
MLrank
01Best alignedFit 96 · Ready now0.96
02QualifiedFit 89 · Nearby0.89
03AvailableFit 83 · Fast response0.83
02 / Decision systems
A recommendation earns trust when the constraints are legible.
My work joins candidate quality, operational eligibility, and changing network conditions into a ranked action that users can understand and improve.
01DefineNeed + outcome
→
02FilterHard constraints
→
03ScoreFit + context
→
04LearnOutcome feedback
30K+
vendor-network context behind operational matching and recommendation work.
“The useful model is the one that understands the decision around the prediction.”
Chapter IIILearning when the world changes
Adaptation
Online meta-learning and feedback loops for systems that cannot assume tomorrow looks like yesterday.
View the researchModel loss / changing dataConcept view
HighLossLow
shift 01shift 02
Static baseline Adaptive modeltime →
03 / Applied ML research
Models should learn from change, not merely survive it.
My doctoral work explored continuous model adaptation using online meta-learning—research that still shapes how I think about monitoring, evaluation, and feedback in production AI.
289%
measured loss-value improvement reported in the research context
View IEEE record ↗
Observe drift
Watch the relationship between current data and learned assumptions.
Adapt selectively
Update what needs to move without forgetting what still works.
Measure the loop
Evaluate the system across time, not just at a frozen benchmark.
Chapter IVResearch depth, production pace
Experience
From adaptive-learning research to AI systems inside a large property-services platform.
Open the timeline
Now
Senior Data Scientist
Lessen · Chicago
Agents, recommendations, semantic knowledge, and technical leadership.
2017 → 2026
04 / Experience
A career built around systems that have to work.
2024 — NowSenior Data ScientistLessen
Designing recommendation, agent, communication, and semantic-knowledge capabilities across a large property-services ecosystem.
- LLM agents
- Recommendation systems
- Evaluation
- Leadership
2022 — 2024Data Scientist IISMS Assist / Lessen
Delivered location-scale scoring releases, ML pipelines, and cloud data workflows for operational decision systems.
- Python
- SQL
- AWS
- Model releases
2022Data Services InternAssurant
Supported data-centric applications and Data as a Service capabilities using Python, SQL, and applied ML patterns.
- Data services
- Machine learning
- Python
2017 — 2022Ph.D. ResearcherUniversity of Illinois Chicago
Researched online meta-learning, deep-learning optimization, and continuous model adaptation for smart-grid applications.
- Meta-learning
- Deep learning
- Optimization
- IEEE
Chapter VIdeas built to travel
Research
Academic work in adaptive intelligence, carried forward into practical AI design.
Read the record
IEEE TNNLS / 2020
Continuous Model Adaptation Using Online Meta-Learning for Smart Grid Application
Jinghang Li et al.Adaptive intelligence ↗
05 / Selected research
Formal research, practical instincts.
Peer-reviewed work, a doctoral thesis, and a continuing interest in systems that learn under real-world change.
Journal article · IEEE TNNLS
Continuous Model Adaptation Using Online Meta-Learning for Smart Grid Application
Adaptive prediction under shifting training and real-time data patterns.
View IEEE record ↗ Doctoral thesis · UICContinuous Model Adaptation Using Online Meta-Learning
Online meta-optimization, deep neural networks, and real-time change.
Read the thesis ↗ Research profilePublications and citations
Google Scholar ↗ Research identityVerified academic record
ORCID ↗
Chicago, IL · 41.8781° N
Afterword / About
I’m Jacky—a scientist who likes messy problems, clear systems, and work that ships.
My path started in industrial engineering and adaptive-learning research. Today I bring that same curiosity to production AI: understanding the system around the model, asking sharper questions, and helping cross-functional teams move from ambiguity to action.
- Current role
- Senior Data Scientist at Lessen
- Core toolkit
- Python, SQL, AWS, Azure, PyTorch, TensorFlow
- Best conversation
- How AI can make a real workflow meaningfully better
A question, an idea, or a complex system?
Ask me
anything.
I’m always glad to compare notes on applied AI, recommendation systems, adaptive learning, or a difficult workflow that deserves a clearer system.
hello@jackyli.phd ↗