DEV Community

Cover image for I built FrankenData with Kiro: spec-first, then I turned Kiro on my own code
harshith kadaru
harshith kadaru

Posted on Edited on AI-assisted

I built FrankenData with Kiro: spec-first, then I turned Kiro on my own code

At my Salesforce internship, and again running events for a student club, I kept doing the same job by hand: making two exports agree. A CRM export and a billing export. An RSVP form and a ticketing list. Same people, different spellings, emails in different casing, duplicates. Every month it came back.

So for Kiroween 2026 I built FrankenData: a lab where you drop in mismatched spreadsheets and stitch them into one living dataset, right in your browser.

Try it: https://kiroween-frankendata.vercel.app
Code: https://github.com/Kadaru242003/kiroween-frankendata

What it does

Load the business demo and you get a 22-row CRM export and a 48-row billing export that never agree.

  • The Autopsy gives every table a 0 to 100 health score and lists exactly what is wrong: duplicates, missing values, stray whitespace, casing variants. One click queues every fix for review.
  • The stitching table proposes every column match and cleanup with a before and after. Rules catch the obvious ones. A bounded AI catches the rest, like client_name and Customer_Name, and gets an "AI found this" badge. Nothing changes until you approve.
  • Ask the Lab answers plain-English questions. It writes the SQL, shows it to you, runs it locally in DuckDB-WASM and charts the answer.
  • Recipes save the whole cleanup. Next month, load the new files, apply the recipe, and every step replays. Anything that does not fit is reported instead of forced.
  • Little Frank, a small patchwork creature, sleeps in the corner until your data arrives and lightning brings him to life. Ask him anything about the lab and he answers in the voice of the creature from Mary Shelley's 1818 novel, fully offline.

The privacy rule that shaped everything: your rows never leave the browser. When AI is on, it sees only column names, types and a few short sample values. An AI Off switch makes the app fully local.

Spec-first, one small batch at a time

I did not start with code. Every feature started as a Kiro spec: requirements, then a design, then a task list. FrankenData ended up with eight specs: file intake, AI stitching, chat SQL, autopsy, recipes, an MCP server, smarter questions, and even one for Little Frank.

Writing requirements first is what kept the AI parts safe. "Rows never leave the browser" and "AI-written SQL must be read-only" became acceptance criteria with tests, not good intentions.

Steering kept Kiro's context small. Four docs are always included (product, tech, structure, security). Five are file-matched, so a change to a React component pulls in the design system and a change to the SQL engine pulls in the data-engine rules. One is manual: a release checklist I pull in before every merge.

Hooks were built for this project's own risks. The one I relied on most reruns the SQL validator's property tests whenever the validator changes. Another rebuilds and tests the MCP server when its source changes.

An MCP server exposes the app's own engine as 8 tools. From Kiro chat I ran an Autopsy on the CRM demo file and got 79/100, the same score the web app shows, because both call the same code.

Then I turned Kiro on my own code

Once the features worked, I spent the last stretch using Kiro to attack FrankenData instead of extend it. That turned out to be the most valuable part of the build.

My security agent found a real bypass. I had written two custom agents early on, a spec reviewer and a security auditor, and barely used them. Before submitting, I pointed the security auditor at my SQL validator. It found that a table listed after a subquery skipped the allow-list check, and that the auto-repair step could send a cell value to the AI inside an error message. Kiro reproduced both with failing tests first (26 bypass cases failed against the old validator), then fixed them. It also added a second line of defense: DuckDB now runs with file and network access disabled and locked, so even a query that slipped past the validator could not read anything outside the browser.

My own test files exposed a join bug. I made a vendor CSV and a purchase spreadsheet by hand and dropped them in. The stitch joined on country instead of email and doubled the rows. Digging in, Kiro found a worse problem: the Join key dropdown never changed the SQL at all. A fixture test failed first, then the fix landed. Next month's recipe replay went from 5 unmatched invoices to 0.

A bug sweep against the production build found four more. I asked Kiro to audit 13 areas and fix only what it could prove with a failing test. It found CSV formula injection in exports, Excel dates shifting by a day in some time zones, leading zeros lost from IDs, and a stitch that failed silently on mixed-type keys. All four are fixed and pinned by tests.

The numbers

  • 8 specs, 10 steering docs, 9 hooks, 2 custom agents, 1 MCP server with 8 tools
  • 1,451 unit and property tests across 214 files, plus 27 Playwright end-to-end tests
  • Every change runs the same gates in GitHub Actions: lint, tests, build, a serverless check, the MCP build and end-to-end tests

What I learned

Specs and steering get you a working app. The agents, property tests and end-to-end tests are what made me trust it. Every bug they found was exactly the kind a real user would have hit first, and none of them showed up while I was clicking around my own demo.

If you try one thing from this post: write the scary requirement down first ("never send rows", "read-only only"), then let Kiro turn it into a test. Then point an agent at your own code and ask it to break it.

Built for #KiroweenHackathon26 in the Frankenstein category. Try it at https://kiroween-frankendata.vercel.app and tell me what breaks.

Top comments (0)