We leavea traceon the future

Measures weekly how often, in what position and in what tone a brand appears in AI answers, and which sources get cited. Built for the Turkish market.

We apply what we recommend to ourselves first: humanizemy.ai, metegpt.com and the evidence-protocol review site mimoaitools.com.

We train a small LoRA on a base model with no instruction tuning and apply it round by round. AI text turns into natural human language, and the numbers are preserved.

A classifier tested on text from unseen generators and on academic writing, which doesn't wrongly flag non-native English writers.

An acoustic classifier trained leak-free on two independent datasets. Trained on one set, it stalled at 0.55 on the other; with more varied sources it rose to 0.95.

Audits every run: every value in the answer must appear in a tool output the agent didn't write itself. Counting and summing over local files must be computed by a command.

A change must either fix a known error or cut tokens without changing any result. A held-out anchor set catches overfitting to the exam.

Data and shell tasks go to Qwen 27B dense, web and code tasks to Qwen 35B MoE. The two models' errors were fully disjoint; the exam rose from 40/47 to 47/47 and tokens fell 10%.

When off-the-shelf open models couldn't pass the gate, we trained our own classifier. False positives on student essays fell from 16.7% to 0.5%.

We brought music's self-similarity matrix to text. LLM text comes out more coherent than human text and adds meaningfully to classic measures.

Do writers build surprise like a musical arc? A well-controlled no. The real signal is in level and uniformity; the measure that survives paraphrase attacks is burstiness.

Are numbers, citations and meaning preserved in a rewritten academic paragraph? We measure sentence by sentence. A citation protection layer cut fabrication by about 30%.

A model trained on one set stalled at 0.55 on the other. Augmentation didn't close the gap; a second, different source gave 0.95 and 0.97 on both sets.

Acoustic, direction, proximity, passive RF and vision layers combine with a noisy-OR. In the scenario, when acoustics missed, RF took over and gave a bearing of northeast 45°.

Çelikkubbe and the startup program, then TÜBİTAK and SAHA, then SSB (Turkey's defense industry agency). Position: a detection and warning sensor. Neutralization stays with authorized integrators.

An assistant that answers the hotels' incoming calls in Turkish. It listens to the caller, records the request and passes a summary of the call to reception.

A Turkish assistant that runs entirely locally: it takes messages, schedules callbacks and appointments, and answers common questions. It gives no legal advice. The recording and summary go to the team.

Texting by default; ask and she sends a selfie or a voice note, and "one more in the same outfit" keeps the outfit. Her surroundings change with her time of day.

Chat, stories, build-your-own character and a voice call scaffold. Verified end to end with real taps in the simulator.

A personality coordinate seeded from the user's choices, plus 24 culture packs. In testing, stereotype drift was 0% and separation between clusters 100%.

Each character's voice is derived from its personality card and it speaks in that voice. The same recipe gives exactly the same result in the app and on the server.

Committee-level real votes from three elections, 180 member personas and a core that uses no LLM. The actual seat results of two past elections (38/66 and 51/68) fell within the model's 90% interval.

A voter simulation built from thousands of personas. As the protocol improved, the Alanya 2024 error fell from 6.7 to 4.1 points. Backtested on seven pairs of national elections.

Sourced setup and operating costs for four Istanbul locations, a three-year model, customer persona simulation and a comparison against deposits and the stock market.

90 days, 2,160 hours, 3,557 headlines. News volume leads BTC volatility by about an hour; sentiment and surprise measures carry no signal.

A paranormal universe spanning 50 years and 30 states: an 833-line canon, characters, a theme song and a Season 1 trailer.

20 to 30 second videos for TikTok, Reels and Shorts. A hook in the first frame, the product in the middle, a real proof screen at the end.

Reference-locked characters, frame-by-frame drawing with Canvas, Blender scenes, a Python compositor and picture-locked sound design.

Scan trash on the ground with the camera and, depending on what it is, a creature appears in augmented reality. Collect them and play with others. Five creatures modeled in 3D, the game is built in Unity.

Chat, stories, build-your-own character and a voice call scaffold. Verified end to end with real taps in the simulator.

Mustafa Mıhcı's iPhone app. An estimated starting level from 14 questions, a personal protocol around sleep, training, diet and stress, and progress tracking. Data stays on the device.