AWS Builder Center

Weekend Challenge: Turn One Annoying Task Into an App: NOMEN

NOMEN reads dropped PDFs and images with Amazon Textract and Amazon Bedrock (Nova Micro), then proposes a reviewable filename built from what's actually inside the document — no accounts, no database, nothing stored.

Sophomore at SRMIST | AI & Cloud Developer | Founder of Rakshak Safety Intelligence

Vision & What the App Does

Every downloads folder has these in it:
scan_0001.pdf
IMG_3821.jpg
Invoice(7).pdf
Each of those files contains the information that should have been its name. The invoice knows who issued it, when, and what its reference number is. The filename knows none of it.
Fixing that by hand is a loop: open the file, read enough of it to identify it, switch back, retype a name, move to the next one. It isn't hard work. It's repetitive context switching, and it scales badly — thirty files is thirty round trips.
NOMEN removes the loop. You drop files in, it reads them, proposes a filename built from what's actually inside the document, lets you correct anything it got wrong, and hands back the original file under the new name.
scan_0001.pdf → 2026-07-12__ACME-Corp__Invoice__INV-4432.pdf
Four steps: DROP → UNDERSTAND → REVIEW → DOWNLOAD. No signup, no library, no dashboard to come back to.

How You Built It

The constraint was set before any code: NOMEN does one job. No accounts, no database, no stored documents, no file manager. That decision shaped the architecture more than anything else.
The frontend is React + Vite. When you drop files, the original bytes stay in the browser — they never leave it. For a PDF, page 1 is rasterized to an image client-side and normalized. That normalized analysis image is the only thing the backend ever receives. Dropped images go through the same normalization path.
The backend is a single Lambda behind a Function URL, doing three things in order:
  1. Textract runs on the analysis image — OCR only. It reads what is on the page.
  1. Nova Micro, via Bedrock, interprets that extracted text into structured metadata: document type, party, date, document/reference number.
  1. A deterministic filename builder assembles and sanitizes those fields into the final name.
That third step is the one worth pointing at. The model returns fields, not filenames. Once the metadata comes back, building the name is ordinary code: field ordering, date formatting, character sanitization, length limits. Nothing probabilistic touches the string that becomes a filename. Same metadata in, same filename out, every time.
The proposal goes back to the browser, where you can edit any name before downloading. Download writes the original bytes, untouched, under the reviewed filename.
The review step is deliberate rather than a fallback. Extraction from a single rasterized page will sometimes read a vendor name oddly or pick the wrong date off a document with several on it, and a tool that renames files silently is a tool you have to audit afterwards. Showing the proposed name next to the original and making it editable means a wrong answer costs one keystroke instead of a hunt through a folder later. It also keeps the failure mode honest — you see what the pipeline understood before anything is written to disk.

AWS Services Used / Architecture Overview

NOMEN architecture diagram: browser to Lambda Function URL to Textract and Bedrock Nova Micro to deterministic filename builder and back

  • AWS Lambda + Lambda Function URL — the entire backend
  • Amazon Textract — OCR
  • Amazon Bedrock / Amazon Nova Micro — semantic extraction
  • Vercel — static frontend hosting
Region:  ap-south-1 . Inference profile:  apac.amazon.nova-micro-v1:0 .
Confirmed against the Lambda's own IAM policy: the execution role grants exactly  textract:DetectDocumentText ,  bedrock:InvokeModel ,  bedrock:Converse , and the three  logs:*  actions CloudWatch needs — nothing else. No S3, no DynamoDB, no API Gateway, no Cognito. AWS Lambda, Amazon Textract, and Amazon Bedrock are the complete list of AWS services this project touches; the frontend is a static build and didn't need an AWS host, so it runs on Vercel instead.
What isn't in that list is as deliberate as what is. No API Gateway — a Function URL is enough for one synchronous POST. No S3 upload pipeline, no DynamoDB, no auth, no RAG, no agent framework. Storage would only add something to secure, and then something to delete.

What You Learned

Separating OCR from semantic extraction was the right split. Textract is good at reading pixels and isn't trying to know what an invoice number means. Nova Micro is good at interpreting text and shouldn't be doing layout analysis on a scan. Each stage does one thing and hands off a clean interface to the next.
Keeping originals client-side simplified everything downstream. Nothing stored means no bucket policy, no lifecycle rule, no retention question, no deletion endpoint. The backend never sees the document — only a flattened image of page 1.
Put the deterministic layer after the probabilistic one. Letting a model emit a finished filename would mean validating a freeform string on the way out. Constraining it to structured fields and building the name in code made the output predictable and the failure modes visible: a missing field is obviously missing, rather than quietly producing a plausible wrong name.
Synchronous processing sets the scope. One request, one response, no job queue — that puts a hard ceiling on per-file work, which is what pushed the design toward analyzing page 1 rather than whole documents.
Bedrock in APAC wants the inference profile ID, not the bare model ID:  apac.amazon.nova-micro-v1:0 .
Two things changed between the plan and what actually shipped. First, region and model: the generic deployment runbook was written against  us-east-1  with the  us.amazon.nova-micro-v1:0  profile, but the Lambda that's actually live runs in  ap-south-1  and calls  apac.amazon.nova-micro-v1:0  instead — Bedrock inference profiles are region-scoped, so the prefix has to match wherever the function actually lives, not wherever the docs were written for. Second, hosting: AWS Amplify was on the list of candidate services during planning, but the frontend that's actually live deploys to Vercel — a static React build didn't need anything AWS-specific, so it went to whichever host was simplest to wire up alongside the Lambda Function URL.

Link to App or Repo

Drop zone — before dropping any files

Processing — Textract and Bedrock analyzing dropped files, renamed results appearing as each completes

Result — all files renamed, ready to download individually or as a zip


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