Self-Directed Product · Career Agent · Case Study 04
Separating source facts, AI interpretation, and generated language
A bilingual EN/ZH job and resume intelligence workflow built around one hard rule: generated conclusions and resume claims must remain traceable to verified candidate evidence.
Knowledge EngineeringSemantic EvaluationProvenance
90+Live Test Rounds
EN/ZHBilingual Evidence
0→1Solo Build
01 / Problem & Context
What had to become true
The challenge
A bilingual EN/ZH job and resume intelligence workflow built around one hard rule: generated conclusions and resume claims must remain traceable to verified candidate evidence.
My role
I owned the problem framing, system/product decisions, implementation direction, validation, and iteration represented in this case study. The emphasis is on the decisions that changed the system—not a feature inventory.
02 / Decisions & Tradeoffs
Architecture follows the constraint
Key decisions
- Separate source material, extracted facts, user-confirmed evidence, AI interpretation, and resume-ready wording.
- Replace brittle substring matching with semantic evidence evaluation.
- Make completion depend on persisted evidence and factuality verification, not only generation success.
- Preserve non-destructive version history and human override.
Working loop
PROBLEM → INVESTIGATE → ARCHITECT → BUILD → VALIDATE → ITERATE
The implementation was treated as a measured loop: diagnose the actual constraint, make the smallest architectural change that resolves it, then verify behavior and quality before moving forward.
03 / System Design & Build
What I built
- Canonical Evidence Library stores facts and provenance as structured data.
- Semantic evaluator judges requirement intent, evidence strength, transferability, and uncertainty across EN/ZH.
- Claim-level verification links generated wording back to supporting evidence.
- Independent ATS, recruiter, and hiring-manager review acts as a separate quality layer.
- FastAPI, Next.js/React, TypeScript, SQLModel/SQLite, OpenAI APIs, structured outputs, and persistent workflow state support the end-to-end system.
04 / Validation & Outcomes
Evidence over claims
Validation
- 90+ logged end-to-end testing rounds against real application workflows.
- Testing exposed invisible wait states, incomplete bilingual output, evidence reconciliation failures, and decision-quality regressions that isolated tests missed.
- Architecture changed in response to production behavior rather than only prompt edits.
What changed
For high-stakes generated language, provenance is product infrastructure. The system needs to know not only what it wants to say, but why it is allowed to say it.
05 / Skills & Positioning
What this project demonstrates
Knowledge Engineering · Semantic Evaluation · ProvenanceThis case sits inside a broader portfolio spanning AI systems and agents, full-stack products and revenue infrastructure, design engineering, and creative AI leadership.
Portfolio Map
Explore the full system of work
Each project is a different proof point. The current project is highlighted so the portfolio reads as one connected capability map rather than a collection of isolated case studies.
