LAiSER
Research Specialist · George Washington University · Jan 2025 – present
Open-source AI tools that extract skills from education and workforce data and align them to shared taxonomies. I built the knowledge graph, the fine-tuned model and the agentic version, and mentored five interns.
0.65 → 0.83
LLM faithfulness after grounding in the graph
18K+
skills unified across four taxonomies, with 11.6K links
5
interns mentored across the graph and an NLP paper
3 days
to collect 100K+ syllabi from 31 Texas universities
The problem
Learners, educators and employers all describe skills differently. A syllabus, a job posting and a national taxonomy can mean the same thing and share no words. LAiSER is a George Washington University research initiative that builds open-source tools to extract skills from that text and line them up.
I joined as a research specialist and my scope grew with the project: from a single model, to the knowledge graph underneath it, to the agentic system on top, to mentoring the interns working on it.
What I built
- The LAiSER knowledge graph. It unifies more than 18,000 skills across ESCO, O*NET, UK Skills and OSN, joined by 11.6K links.
- SkillPhi-R. A 2.7B Phi-2 model fine-tuned for skill extraction with LoRA and 4-bit quantization, answering in 10 to 15 seconds. It won the GWU open-source award, and I presented it to about 60 people.
- LAiSER 2.0. An agentic system on an MCP server. Skill, knowledge and task extraction are exposed as callable tools over a Neo4j graph of job, skill, knowledge and task relationships.
- The package. LAiSER ships as a Python package and API used by students and university professionals.
Decisions
decision 01
Fix the grounding, not the model
- Situation
- The LLM extracting skills kept returning skills the source text did not support. The obvious fix was to change or retrain the model.
- Task
- Reduce made-up answers.
- Action
- I kept the same model and grounded its answers in the knowledge graph, so it could only draw on skills that exist there.
- Result
- Faithfulness, how far the answers are supported by the source text, rose from 0.65 to 0.83 with no retraining. The change was in what the model could draw on, not in the model.
decision 02
Thresholds that move with the data
- Situation
- The graph links skills across four taxonomies written by four different bodies. They do not match at one consistent level of similarity, so a single fixed cut-off produced links that needed manual checking.
- Task
- Replace the single cut-off with one that fits the skills being compared.
- Action
- I used dynamic thresholding with automated elbow detection. For each skill, the candidate matches are ranked by similarity score and the cut-off is placed at the largest drop between consecutive scores, with no human reading the curve.
- Result
- Each skill gets its own cut-off, and fewer links had to be checked by hand.
decision 03
Skip the agent when a query will do
- Situation
- In LAiSER 2.0 every request went through the full agent loop, which costs tokens and time. Many questions are plain look-ups against the graph.
- Task
- Keep simple questions cheap.
- Action
- When a question maps onto the graph, it is answered with one direct graph query (Cypher, Neo4j's query language). The full agent loop, where the model checks and revises its own answer, runs only when it does not.
- Result
- Plain look-ups no longer pay for the agent loop, in tokens or in time.
decision 04
Test the graph like code
- Situation
- A bad merge can drop part of the knowledge graph without breaking any code, so normal tests would still pass.
- Task
- Catch a damaged graph before it reaches the people using the package.
- Action
- I added GitHub Actions tests that gate every merge to main, including checks on the number of nodes in the graph.
- Result
- A merge that drops part of the graph now fails in CI.
Mentoring the interns
I mentored five interns: three on the knowledge graph and two on an NLP paper about phantom vacancies, which is in review. I assigned work by each person’s strengths and reviewed code before it merged to main.
One piece of that work was data collection. We needed more than 100,000 syllabi from 31 Texas universities, and an EC2 scraper brought the collection down to three days.