I Got Tired of Fixing Dirty JSON by Hand, So I Built Yellorn

Sat, 02 May 2026 00:00:00 +0000

TL;DR for the folks who are already tired

The problem: Data from LLMs, Python logs, Excel exports, and vendor webhooks does not break in elegant, academically interesting ways. It breaks like raccoons got into your clipboard.
The fix: Yellorn — a Data Doctor that patches busted JSON/XML/YAML/CSV, mocks webhooks on Cloudflare Edge, sends HTTP requests with saved templates, and converts formats in one editor.
The useful bits: 29 auto-fix passes. Every fix is shown, not silently prayed over. Webhooks run on HTTPS. Publishing requires login. The webhook URL is public. The inspection page is owner-only.
The verdict: Does not replace Postman. Does not replace your backend. Does not replace having actual discipline with your data contracts. But it does replace seven suspicious formatter tabs — three of which look like they were designed by a casino affiliate marketer having a spiritual crisis. That counts as civilization.

The Developer Janitor Tax

Last Wednesday, an LLM handed me “valid JSON.”

It was wrapped in Markdown fences, had a friendly intro sentence, and included a trailing comma like it was leaving a tip.

jq hated it. I hated it. The coffee was already cold.

Four minutes later, I had paid the developer janitor tax again.

You know the tax. It is not the hard work. Not architecture. Not distributed systems. Not debugging a beautiful race condition that makes you feel like a detective in a Scandinavian crime drama.

It is worse.

It is removing backticks from LLM output. It is replacing smart quotes from a CSV someone exported from Excel. It is figuring out whether the delimiter is a comma, semicolon, tab, or whatever regional setting decided to wake up and choose violence that morning. It is copying a webhook URL into a vendor dashboard, forgetting to hit Save, then briefly considering farming.

Same morning: a coworker sent a CSV over Slack. Excel on Windows had added a BOM, switched to semicolons, and used typographer quotes. PM needed confirmation it matched the vendor webhook payload — from a vendor whose callback I had not received because ngrok had changed its URL four minutes ago.

Nobody puts this on the sprint board. Nobody estimates:

1
2
3

Fix Claude's emotionally supportive JSON: 0.5 story points
Clean Excel's little punctuation crime scene: 0.5 story points
Rebuild local webhook tunnel because ngrok changed its mind: 1 story point

But the tax gets collected anyway. Five minutes here. Four minutes there. A tiny slice of focus gone. Just enough friction to break flow and make you stare at the monitor like maybe opening a sandwich shop was the correct career path all along.

Yellorn refunds that tax. Not with a manifesto. With a textbox, a few buttons, and a refusal to pretend dirty data is some exotic edge case.

Dirty data is not an edge case. Dirty data is weather.

Data Doctor

JSON.parse is strict. It should be. JSON was built for machines talking to machines, not for negotiating with a blob of text that just took a group tour through Slack, Claude, Excel, Notion, and someone’s clipboard history.

Yellorn starts strict. If the input is valid, great. No drama. If it fails, Yellorn runs best-effort recovery through 29 auto-fix passes across JSON, XML, YAML, and CSV. This is not magic. It is janitorial work, automated.

The things it actually handles:

{
  "name": "Yellorn",
  "works": true,
}

That is not JSON. That is JSON wearing a Markdown hoodie. Yellorn strips the fence, removes the inline explanation, fixes the trailing comma, and tells you what it did.

`1`	`{'ok': True, 'price': Decimal('3.14'), 'created_at': Timestamp('2026-05-02')}`

Python logs pretending to be JSON. Yellorn normalizes the types.

1
2
3

id;name;status
1;"Alice";active
2;"Bob";failed

Excel CSV: BOM, typographer quotes, semicolon delimiter, all at once. In total: 15 passes for JSON, 5 for XML, 6 for YAML, 3 for CSV. Enough to clean up most daily clipboard sludge.

The important part: Yellorn does not quietly hallucinate your data into being “correct.” Every fix is visible — fence stripped, quote normalized, bracket closed, comment removed. You get the receipt. If the data is wrong in a business-logic way, it is still wrong. A tool should not develop a savior complex about your schema. It should just stop commas from ruining your morning.

Mock Webhooks

You need to test a Stripe callback. Or a GitHub hook. Or something called “temporary-integration-v2-final” that has somehow been in production for 14 months.

You know the ritual. Start local server. Open tunnel. Copy URL. Paste into dashboard. Hit test. Nothing. Check logs. Nothing. Check tunnel. Dead. Restart tunnel. Copy new URL. Paste new URL. Hit test again. Dashboard kept the old URL because you forgot to hit Save.

You stare at the screen and wonder how humanity put people on the Moon but still cannot make webhook debugging feel like less than reading tea leaves with response headers.

Yellorn makes the dumb part shorter. Paste a payload into the editor. Hit Publish. You get an HTTPS endpoint:

https://yellorn.com/api/webhook/<slug>

Runs on Cloudflare Edge. Publishing requires login — storage is not a magical public wishing well even if the internet keeps acting like one.

The webhook URL is a public capability URL: anyone with the slug can call it. That is the whole point. But the inspection page is owner-only. Outsiders can send requests in. They cannot read your logs. This sounds obvious, but the history of web security has taught us that “obvious” is a premium feature that ships late.

The portal shows: method, query params, headers, body, response status, timestamps. You can simulate response codes — 200, 401, 429, 500 — to test success paths, auth failures, rate limits, and sad little server explosions without building a fake backend each time.

Free tier: 5 active webhooks, 512 KB payload, 24-hour auto-expire, live request log. Pro and Team raise the limits.

Creating an entire Express app just to return res.json(payload) is one of the quieter indignities of software development. Yellorn replaces that ritual with a button. Not a production server. A field airstrip, but it has HTTPS and runway lights.

Request Sender

Webhooks answer the inbound question: where does the vendor send the callback? But integration debugging has another half: what am I sending out, with which headers, which auth, and what came back last time?

Yellorn includes a Request Sender in the browser. GET, POST, PUT, PATCH, DELETE, HEAD, OPTIONS. Query params, headers, Bearer/Basic/custom auth, JSON/text/form-urlencoded body. Click send.

Opening Postman to send one quick request is like renting a forklift to move a sandwich. Postman is great — it has workflows, collections, environments, collaboration, monitors, probably a small government inside the UI. Yellorn is not trying to be that. It is the tab you actually needed: send this request, save this template, show me what happened, do not make this a lifestyle.

Requests can be saved as templates: name, method, URL, headers, auth, body. Each dispatch logs history — status, timing, response headers, body (truncated at 1 MB). When a request fails, you have something to inspect instead of relying on memory — a storage engine with no ACID guarantees and aggressive garbage collection after two beers.

The two sides connect. From a webhook portal, you can open that URL in Request Sender and fire a test request back into it. The sender includes guardrails: SSRF protection, private hostname blocking, forbidden header stripping, 15-second timeout, loop guard via X-Yellorn-Internal-Dispatch. The goal is easier debugging, not “use my browser to scan your internal network” — a business model best avoided if you enjoy sleeping.

Format Conversion Without the Ritual Humiliation

Yellorn understands JSON, XML, YAML, CSV, and Python literals. Auto-detection runs as text changes — paste CSV, it knows; paste XML, it knows. Pin the format manually and it listens. A machine that listens to humans remains a charming feature in 2026.

Format is politics. Backend lives in Python dicts. Frontend lives in JSON. Ops sends YAML. Business sends CSV. Some legacy service still spits XML because it was written when people genuinely believed SOAP was the future. Nobody is necessarily wrong. The person trapped in the middle is you.

Conversion goes through a canonical parsed value, then serializes to the target format. The Python emitter outputs datetime.fromisoformat(...) for ISO date strings, instead of making you personally relitigate the month/day ordering conflict that has divided civilizations. Conversions that may lose information are labeled lossy — forcing deeply nested data into CSV without a warning is “convenient” the same way deleting production data is “fast.”

CSV is not really a format. CSV is a family argument with commas. Yellorn cannot make the family healthy, but it can keep the argument readable.

Looking at Data Without Melting Your Brain

Raw nested JSON is fine until it is six levels deep and you are counting braces with your eyes like some kind of Victorian punishment.

Tree view lets you drill into nested objects without losing your place or your dignity. Table view works for CSV and row-shaped data — especially those Excel exports with the suspicious delimiter of the week. Graph view shows parent-child structure instead of making you mentally parse bracket soup.

Search supports plain text and JMESPath. Instead of searching for failed and manually digging through 200 matches like a raccoon in a dumpster, you write:

`1`	`items[?status=='failed'].id`

And get the exact IDs you need. JMESPath is lazy-loaded only when used, because shipping extra runtime for a hypothetical future moment is how frontend bundles become haunted houses.

Small Mercies

Some features are not glamorous. They just prevent daily paper cuts.

Compare Mode: Pick two tabs, open a read-only Monaco diff editor. JSON keys are sorted before diffing, so reordered keys do not trick you into thinking your API underwent a constitutional crisis. Use it for staging vs production responses, old fixture vs new fixture, two files someone swears are “basically the same.” “Basically the same” has caused more bugs than some programming languages.

Cloud Backup: Log in, back up tabs, restore on another machine. API keys do not sync. Some things should not travel.

Format ladder: Smart (heavy rescue), Simple (parse and pretty-print), Safe (layout cleanup without reparsing), Compress (minify). Each has a purpose. You could collapse all of that into one “Format” button and pray. That is also how incident retrospectives are born.

What It Is Not

Yellorn is not trying to become your API operating system. It is not trying to eat your team workflow. It is not asking you to migrate your life into a workspace.

Request Sender is for quick sends and saved templates, not a full collaboration suite. Yellorn does not replace integration tests, schema agreements, or organizational discipline. If your company uses final_final_v3_really_final.csv as a production contract, software can only do so much before therapy gets involved.

Yellorn also does not make bad data semantically correct. It can fix the commas. It cannot fix the meeting where nobody agreed what the payload meant.

But in the very real minutes of the job — cleaning dirty JSON, inspecting webhook payloads, replaying requests, converting CSV to JSON, checking nested paths with JMESPath — it does the work. Fast. In one place. Without the haunted carnival of online formatters.

The Free tier is enough for personal debugging: 5 active webhooks, 24-hour auto-expire, live request logs, Request Sender, Data Doctor, format converter, compare mode, tree/table/graph views. Pro and Team on Polar if you need bigger payloads, more webhooks, or team workflows. No “contact sales.” We are not at that stage of life yet.

Try it at yellorn.com.

Because next Wednesday, an LLM will hand you JSON wrapped in a TED Talk and jq will start screaming. You know it. I know it. Claude knows it too, and it is still going to wrap the JSON in a Markdown fence.

Not because it is smart. Because it is only slightly less dumb than yesterday.

Yaml on HoangYell