Tibor Czekay | Independent software engineer

Built by me, running now

My agentic super system.

How I stay ahead in a world of daily change. Seven agents I wrote and run myself, on a three-hour loop, reading what people are actually stuck with so I do not have to. Topic to App is built to destroy ideas rather than produce them, and to say exactly why each one died.

7agentsThree hunt, one collects, one judges, one reports, one holds the brief.
8×a dayA full cycle every three hours, unattended, night included.
6job familiesTwo per cycle, never the same pair twice in a row.
27kill classesA dead shelf is written back as a shape, not as one dead idea.

An agent is a small program that reads something, decides one thing, and passes the result on. Seven of them run Topic to App on a three-hour loop, and they only ever talk to each other in messages.

Three of them hunt live sites for somebody describing a repeated moment, a workaround, or a tool that went bad. A fourth sits outside the hunt and dumps thirty days of raw material in bulk. None of them may have an opinion. One agent, and only that one, decides whether a complaint is worth building. One more decides whether it is worth my attention. Splitting the roles is the whole trick: the agents that find things can never talk themselves into liking them.

Scouts log. Three hunt live, one collects in bulk, and not one of them judges.
One critic judges. Ship, maybe, or kill. Nobody else gets a vote.
One manager reaches me. The other six have no way to contact a human.
A kill teaches the brief. Rejections go back as a job class, not as one dead idea, so the same shape is not proposed twice.

topic-to-app · cycle replay

Running

Briefthe spec
Redditscout
Xscout
Blueskyscout
Pasterycollector
Criticthe only judge
Managerthe only way out

transcript

BriefCurrent spec: solo-shippable, on-device, no ongoing server cost.

0 judged0 killed0 shipped

The ledger

What it has refused so far.

Hunting · next cycle every 3h

A filter is judged by what it turns down. Since 22 August 2026, 19:00 this one has taken 139 candidates and passed none of them through. That is a count of refusals, not proof that the shelf is full, and the panel at the bottom is why I will not call it proof yet.

22 Aug running since Switched on 22 August 2026 at 19:00, Europe/Berlin. Nobody restarts it in the morning.
139 judged Candidates carried all the way to a verdict with a written reason.
139 killed Every one so far, each filed against a class so the shape does not return.
20 in calibration Control apps, selected against constraints written down before collection began.

verdicts by source

Live huntx.com, reddit.com, bsky.app. Three scouts, eight cycles a day

24 kill0 maybe0 ship
30-day dumpPastery, weekday mornings, collect only, deduped by URL

115 kill0 maybe0 ship

22 Aug3-hour loop, one scout
Split6-hour, three scouts, because the 3-hour reports got noisy and expensive
26 Aug3-hour again, night included, a ping every cycle even when empty

Where this number stops being evidence

Zero hits across 139 rows has two explanations that look identical from inside the system: the shelf really is full, or a three-way verdict whose middle option has never once fired is simply broken. A scoreboard cannot tell those apart, so I am not going to let it try.

The control set is built to separate them: my eight shipped apps plus twelve external indie iPhone apps, drawn from Apple Design Awards, MacStories, App Store features and Reddit indie threads, filtered on constraints written down before collection and sampled every second app by first release date. Control rows never enter the live lists. The critic has not been retuned. Until that runs, the number above is a count and nothing more.

Selected work

Built. Measured. Bounded.

Cyan is a number that was actually measured. Amber is where that number stops being evidence.

Mac · local analytics · 2026

Building

Sonventory

Local-first analytics for TikTok, Instagram, and YouTube, in one traceable report on your own Mac.

The hard part: Three platforms expose different data, media formats, rate limits, privacy rules, and portability archives.
What I built: A sandboxed Mac pipeline for ingestion, recognition, sentiment, statistics, and export. A small cloud broker handles tokens and stores no user content.

SwiftSwiftUIShazamKitCloudflare

Measured0.73sCold path down from 74.5s on the same 7,559 comments, bucket totals unchanged.

Boundary33 casesEnough sentiment corpus for regression checks, too small for a broad accuracy claim.

iPhone · privacy · CoreML

Live

Trutz

On-device image privacy, audited on the bytes it actually exports.

The hard part: “Protected” is not one measurable state. Metadata, visible content, model effects, and transformed exports each need their own check.
What I built: A local search and evaluation loop with optimised targets, held-out objectives, a random control, and verification of the finished file.

Measured5 of 5Metadata block families removed from a crafted JPEG, GPS through Photoshop, with 3840 × 2160 preserved.

BoundaryNot universalResults hold inside the surrogate set and transforms shown, not against every third-party model.

Open the Trutz page

iPhone · tracking · sensor fusion

Live

Antizipeer

Real-time object tracking and motion forecasting from camera, depth, and phone movement.

The hard part: A camera sees pixels, not motion. Phone rotation moves the whole scene, detections drop out, and useful depth depends on the hardware.
What I built: CoreML detection, Vision tracking, CoreMotion correction, and a six-state Kalman filter over position, velocity, and acceleration.

Measured19Focused invariant tests across projection, convergence, clamping, buffering, orientation, and guidance.

BoundaryNo MAENo ground-truth trajectory corpus and no forecast error figure. A research instrument, not a safety product.

Open the Antizipeer page

Method

Speed, with receipts.

01Find the risk
What can go wrong, who pays for it, and which data must never leave the device.
02Move one signal
Get one real input through the model, the interface, and the output before widening anything.
03Test the claim
Keep the baseline, label the examples, test the invariants, rerun the same case.
04Publish the limit
Say what was measured and where it stops. The boundary ships with the product.

Background

Independent builder, enterprise operator.

One side keeps me fast. The other keeps me exact. Most people have one of them.

On my own

12 products shipped solo

Nine live on the App Store, two waiting on their listings, one Mac app in development. Research, build, design, listing, and support, all of it mine.

81 downloads in the last 90 days. These were built to prove I can carry a product from idea to release on my own, not to acquire users. There is no marketing spend and no growth work behind any of them, and I would rather print the number than let you assume a bigger one.

In regulated banking

3,100 workers under access control

Joiner and leaver processes and access administration for around 3,100 contingent workers in a BaFin-regulated environment, on assignment through Apleona: who may reach what, recertification of accounts and entitlements, and the audits that check all of it against the record with external examiners in the room. Nothing there ships on assurance alone, which is where the habit on this page comes from.

Trained

AI & Data Engineering, 2026

Master-level certification, IU International University. BA Game Design, Macromedia. Google AI Certificate Series, IBM Cybersecurity Fundamentals and Architecture.

SwiftSwiftUICoreMLVisionPythonTypeScriptCloudflareGerman and English

The app shelf

More shipped than said.

Nine live on the App Store, two waiting on their listings, one Mac app in the workshop. Every one has its own page with what it does, what it needs, and where it stops.