Jason Liu Portfolio
Developer Experience Engineer on the Codex team at OpenAI
Creator of Instructor, a Python library for structured LLM outputs (13.9k GitHub stars) that OpenAI named among the inspirations for its Structured Outputs API. Now a Developer Experience Engineer on OpenAI's Codex team. The consulting practice and cohort courses are closed, and the site is now writing, investments and an archive.
Location
Somewhere on planet Earth, most likely near a laptop.
Companies
Credentials
- University of Waterloo — Bachelor of Mathematics, Computational Mathematics & Statistics
Skills
Notable
- Creator of Instructor, structured outputs for LLMs (13.9k GitHub stars, 1.3k forks)
- Instructor named in OpenAI's Structured Outputs launch post as one of five open-source libraries that inspired the feature
- Former consulting clients named on his site: Zapier, HubSpot, Weights & Biases, Modal Labs, Timescale, Pydantic
What makes it work
A breakdown of the choices that make this portfolio stand out.
A library whose impact is independently verifiable
OpenAI's Structured Outputs announcement lists instructor alongside outlines, jsonformer, guidance and lark as inspiration. That is an external, checkable validation of the work's real-world impact, far stronger than a self-reported usage claim.
Open-source work that fed a consulting practice, now closed
Instructor was the credibility base for a solo consulting practice (Zapier, HubSpot, Weights & Biases, Modal Labs, Timescale, Pydantic) and for cohort courses on RAG. The site now says plainly that those offerings are closed and keeps the archive readable. It retires a business without deleting the proof of it.
Willingness to publish both technical and personal reflection
Mixing technical writeups with more personal essays signals a real, specific person behind the work rather than a faceless open-source maintainer account.
What AI Engineers can take from this
Specific, actionable tips to apply to your own portfolio — no generic advice.
If your open-source work influences a larger platform or company, note that connection explicitly and specifically — it's stronger proof than describing your own impact.
Use a well-scoped open-source project as the credibility base for a paid consulting or training offer built directly on top of it.
Don't be afraid to publish some personal, reflective writing alongside technical content — it makes the work read as a real person's practice, not an anonymous feed.
Case studies
The homepage says the consulting and courses are closed, then leaves every playbook up. Jason Liu retires a business the way a well-run library deprecates an API: announced, documented and still readable.
Systematically Improving Your RAG
The step-by-step runbook he used with consulting clients to improve RAG systems. Generate synthetic questions per chunk to baseline retrieval recall, combine full-text and vector search, extract metadata, collect user feedback, cluster failing queries by topic and capability, and keep testing as real data arrives.
It is measurement-first RAG work, starting with recall baselines before touching generation, which is the discipline teams need before scaling an AI feature.
Synthetic question generation per text chunk to measure retrieval precision and recall
Real comparison of full-text vs. embedding search, with numbers
Clusters failing queries by topic and capability to decide what to build next
Projects
Instructor
A library for getting reliable, validated JSON from any LLM. You define a Pydantic model, pass it as response_model, and get back a typed object, with validation errors and retries handled for you. It works across providers through instructor.from_provider(...).
Schema-first LLM output is a core production-AI problem. The library solves it with a small API over Pydantic, and a major platform later shipped the same idea natively.
Pydantic models as the contract: type safety, validation and IDE support for LLM outputs
Automatic retries on validation failure instead of hand-rolled JSON parsing
Cited by OpenAI as an inspiration for its Structured Outputs feature
More AI Engineer portfolios
Ready to build your portfolio?
Follow the AI Engineer roadmap — skills, projects, and timeline to get hired.