AWS Builder Center
Building SuguPro: JIS Resume, English CV & AI Polish

Building SuguPro: JIS Resume, English CV & AI Polish

How I built a browser app that generates a JIS resume, career history, English CV, and cover letter, with AI-assisted editing on Bedrock AgentCore Runtime.

Cloud Engineer
I ended up building four different job-application documents at once: a JIS-format Japanese resume, a Japanese career history sheet, an English resume, and a cover letter for the overseas roles I was also considering. Each one wants slightly different fields, and re-typing the same certification dates into four documents got old fast.
So I built SuguPro, a browser app that generates all four from one set of inputs, previews them live, and exports to Excel or PDF. Everything lives in the browser's localStorage. Nothing touches a server except one optional feature, and that feature is the interesting part of this post: AI-assisted editing running on Amazon Bedrock AgentCore Runtime.
The resume and career history tabs follow the JIS format Japanese employers expect, ruled tables and all, with a blank photo box sized for pasting a printed photo by hand rather than uploading one. Rows for education, work history, and licenses are unbounded, since anyone who has changed jobs a few times needs more than three lines. Both documents export to Excel through exceljs and to PDF through html2canvas-pro plus jsPDF, with a toggle between A4 and B5 paper sizes for the resume.


Licenses and certifications turned out to be the fiddliest part of the form. I built a picker with five categories: AWS certifications, other vendor certifications, Japan's IT engineer exams, general national qualifications, and driver's licenses. Checking a box drops the certification name straight into the list below, and a free-text field stays available for anything not covered. Getting the AWS names right took an actual pass through the AWS certification pages rather than working from memory. SysOps Administrator – Associate has been superseded by CloudOps Engineer – Associate under exam code SOA-C03, and Generative AI Developer – Professional exists as its own credential now.

The English resume and cover letter live on separate tabs with a genuinely different shape, since a Japanese-style resume expects a photo, birth date, and gender, and hiring managers writing job ads for overseas roles generally do not want to see any of that. Contact, Professional Summary, Work Experience, Education, Skills, and Languages cover the resume; certifications pull from the same list the Japanese resume tab already has, since there was no reason to type them twice. The cover letter reuses the sender's name and contact info from that same tab and turns three paragraphs (opening, body, closing) into a standard business letter.



Wiring an AgentCore Runtime agent behind an "AI Polish" button

Motivation statements and self-PR paragraphs are the fields people write worst about themselves. I added a Strands Agent running on Amazon Bedrock AgentCore Runtime, invoked through a single "AI Polish" button next to each of those fields. Clicking it sends only that field's text to the agent and gets back a revised version plus a one-line explanation of what changed. Nothing gets rewritten automatically; the suggestion sits there until the user clicks "Apply."
The agent itself is small. It's a Strands  Agent  backed by Claude Sonnet 4.6, with a system prompt that changes depending on which field is being polished, and instructions not to invent facts that aren't in the original text.

Deployment produced two real bugs worth writing down.  agentcore deploy  picked  direct_code_deploy  instead of a container build, because my local Python was 3.14 and outside the runtime's supported range. That deployment mode ignores the Dockerfile entirely and just runs  python agent.py  with no arguments, but my entrypoint script was written to expect  python agent.py serve . Every invocation fell through to an interactive CLI loop and hung until AgentCore Runtime timed out:
Invocation failed: An error occurred (RuntimeClientError) when calling the
InvokeAgentRuntime operation: Runtime initialization time exceeded. Please make
sure that initialization completes in 30s.
CloudWatch Logs showed the interactive prompt sitting there waiting for stdin. The fix was a one-line change to the entrypoint condition so it starts the server whether or not an argument was passed.
The second bug was an IAM region mismatch. Claude Sonnet 4.6 is only reachable through a us-east-1 cross-region inference profile, but the execution role AgentCore Runtime generated was scoped to ap-northeast-1, the region I'd deployed the runtime itself into:
botocore.errorfactory.AccessDeniedException: An error occurred (AccessDeniedException)
when calling the ConverseStream operation: ... is not authorized to perform:
bedrock:InvokeModelWithResponseStream on resource: arn:aws:bedrock:us-east-1:...
Adding a statement to the role's inline policy for the us-east-1 inference profile resource fixed it. It's an easy thing to miss: the runtime's home region and the model's inference region aren't the same thing, and nothing forces them to match.


Two more surprises after deploying to Vercel

Vercel talks to the AgentCore Runtime through OIDC Federation rather than a stored access key, so there's no static AWS credential sitting in an environment variable. The first implementation read  process.env.VERCEL_OIDC_TOKEN  directly to build the web identity credentials, and every request failed in production with a blunt error:
Could not load credentials from any providers
The token Vercel issues at build time and the token a running Function actually needs are not guaranteed to be the same thing; Vercel documents that the live token arrives via an  x-vercel-oidc-token  request header, and reading  process.env  directly is the wrong layer to read it from. Switching to the official  @vercel/oidc-aws-credentials-provider  package, which reads the header correctly, fixed it in one line.
The other surprise was a React hydration error in production only:
Minified React error #418
The resume's "created" date was set with  new Date()  at render time, computed once on the Vercel Function (running in Virginia) and once again on the client (running in Japan time). For a few hours around midnight JST, "today" differs between the two, so the server-rendered HTML and the client's first render disagreed and React discarded the tree. The fix was to leave the field empty on first render everywhere, then fill in today's date from inside a  useEffect  that only runs after the client has mounted, so server and client always agree on the very first paint.

What it looks like running

I ran my own job history through it end to end, ten-plus history rows and all, and the ruled-table layout held up without any manual adjustment. AI Polish turns out to be more conservative than I expected. It tightens phrasing and swaps casual words for formal ones, but it won't invent the specific achievement a vague sentence is missing, which is exactly what the system prompt asks it to do.
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article