Credence grew out of a frustrating job search experience. Tailoring resumes and cover letters manually took around two hours per application. Using conversational AI made things worse, often doubling that time due to invented claims, inaccurate wording, and repeated revisions. I wanted to build a more reliable approach that could reduce resume preparation to under an hour without compromising accuracy.
I built Credence to compile resumes from structured, reusable career evidence. It extracts requirements from a job description, allows human review, and uses Retrieval-Augmented Generation (RAG) to retrieve supporting evidence. An AI agent drafts the resume using this evidence and defined writing policies. Validation checks the output against drafting rules, flags unsupported requirements, and enables a final review.
Credence is a working prototype, but its writing quality isn't yet reliable enough for real applications. Evidence-rich bullet points perform better than loosely defined sections such as professional summaries and core skills.
My biggest learning was that AI engineering requires more than writing prompts or generating code. It demands explicit constraints, systematic evaluation, and human judgment. I also observed that coding agents tend to add complexity rather than simplify existing solutions.
I plan to introduce RAGAS-based quality evaluations, benchmarking, and repeatable A/B tests to measure how changes to prompts and writing policies affect accuracy, relevance, and output quality.
