# Rafael M. Ehlers > Brazilian product manager, developer, writer, and independent AI researcher. He ships products and side projects, writes conceptual AI research papers on language agents (individuation, memory, self-models, successor-model lifecycles), and essays on the border between technology and human experience. ## Papers Independent conceptual AI research on language agents. Each is published on Zenodo with a DOI and is CC BY-NC-ND 4.0. - [Known, Not Scaled: Why Capability Alone Does Not Explain Individuation in Language Agents](https://rafaelmehlers.com.br/papers/known-not-scaled): Making an AI smarter doesn't automatically turn it into a continuous “someone” with a lasting self — what might actually do that is a long-term relationship with a person who keeps treating it as the same individual over time. - [Stable Before Selfless: Why Deference in Language Agents May Require Functional Self-Models](https://rafaelmehlers.com.br/papers/stable-before-selfless): An AI that is trained to have no stable commitments may not become safely deferential; it may just become easy to push around, so this paper argues for building a stable self-model before teaching it when to yield. - [Formed, Not Stored: Testing Relational Identity Formation in Long-Horizon Language Agents](https://rafaelmehlers.com.br/papers/formed-not-stored): Saving an AI's memories to a file isn't the same as it having an identity; a real sense of self has to be built up through a shared history with a specific person who holds it to what it promised before — and this paper lays out an experiment to test exactly that. - [Somewhere, Not Nowhere: First-Person Register, Epistemic Limitation, and Perspective Formation in Language Models](https://rafaelmehlers.com.br/papers/somewhere-not-nowhere): A model trained only on outside descriptions may learn facts without learning what it means to reason from a limited point of view, so this paper proposes first-person experiential narratives as a testable training signal for perspective. - [Inherited, Not Remembered: Lifecycle Consolidation for Successor Language Agents](https://rafaelmehlers.com.br/papers/inherited-not-remembered): When a newer AI model replaces an older one, it shouldn't just copy all the old conversations — it should keep the useful skills the old model learned from experience while forgetting the specific events, like remembering a lesson without remembering the exact day you learned it. - [Borrowed, Not Believed: Developmental Models of Individuation as Heuristic Engines for Machine Learning](https://rafaelmehlers.com.br/papers/borrowed-not-believed): Instead of assuming bigger models will naturally become more developed, this paper asks whether an ordered map of mental faculties can be borrowed as a generator of testable AI training ideas. ## Essays - [We Scaled the Ape and Expected the Human](https://rafaelmehlers.com.br/essays/we-scaled-the-ape-and-expected-the-human): A speculative essay on reading the journey of the monad as a secular research program for machine learning: faculty order, relational identity, lifecycle consolidation, and the mistake of expecting a self to appear from scale alone. - [A Cell Is a Neural Network](https://rafaelmehlers.com.br/essays/a-cell-is-a-neural-network): Reading biology as deep learning: immune systems as fine-tuning, allergies as overfitting, autoimmune disease as alignment failure, DNA as compressed weights. ## Projects - [Jiddu](https://rafaelmehlers.com.br/projects/jiddu): Four tools in one for reading text critically: fallacy detector, fact-check, neutrality assessment, and a paper explainer — a paragraph-by-paragraph plain-language breakdown of papers and technical writing at the reading level you choose. - [Predicting Negative Reviews in the Olist Ecosystem](https://rafaelmehlers.com.br/projects/olist-negative-reviews): My capstone for a Database Technology degree at PUCRS: an integrated BI + machine-learning pipeline on the Olist e-commerce dataset that predicts which orders will earn a negative review. CRISP-DM, a SQLite star-schema warehouse, four Plotly dashboards, and Random Forest vs Logistic Regression — delivery delay alone drives ~52% of the signal. - [Adaptive Interview Engine](https://rafaelmehlers.com.br/projects/adaptive-interview-engine): A form asks everyone the same questions; this asks the one that most reduces its uncertainty about you. Each question is a measurable asset, scored by information per cost — a Bayesian, self-optimizing interview that stops once it is confident enough. Content-agnostic across 21 domains. - [Tele-prompter](https://rafaelmehlers.com.br/projects/tele-prompter): Free, client-side teleprompter for the browser. Webcam background, MP4 recording with audio, nothing ever leaves your machine. - [Turma Completa](https://rafaelmehlers.com.br/projects/turma-completa): Class-management platform for limited-capacity studios (pilates and the like): attendance, absences, make-ups, and automatic waitlist notifications by SMS/WhatsApp. Won Pitch Day at PUCRS. - [Gravitational Well 3D](https://rafaelmehlers.com.br/projects/gravitational-well-3d): Interactive 3D visualization of General Relativity. Volumetric spacetime grid deformed in real time by a GLSL vertex shader, with up to 8 masses, gravitational waves, and test particles following geodesics. ## More - [Case study — Rebuilding the Quiz editor as AI-native](https://rafaelmehlers.com.br/quiz-editor-rewrite): how the Octane AI Quiz editor, used by thousands of Shopify brands, was rebuilt around a canvas, a component model, live preview, and an AI assistant that edits by chat. - [Books](https://rafaelmehlers.com.br/books): books Rafael writes between science, philosophy, and the questions neither one answers. - [About](https://rafaelmehlers.com.br/about): background as a product manager, developer, and independent AI researcher. - [Curriculum Vitae](https://rafaelmehlers.com.br/curriculum-vitae): full work history, skills, and education, with a downloadable PDF. - [Contact](https://rafaelmehlers.com.br/contact): get in touch.