Applied AI • DevNet • Computational Materials • Systems Thinking
Building experimental architectures at the intersection of AI, physics, and infrastructure.
Maximiliano Rodrigo Speranza — Independent Researcher — Buenos Aires, Argentina 📧 maximiliano.speranza@gmail.com · 🆔 ORCID 0009-0005-0413-8554
🇺🇸 English
I am an independent AI researcher, and for over 20 years I have been solving problems where hardware, software and the physical world meet. I study how to make a language model remember what it was told, correct it when it changes and say "I don't know" about what it never heard, through preregistered experiments and published results, including the ones that fail. By day I run Westo, where I bring the same way of working to construction and industrial installations, and I build my own quantitative trading systems. One simple idea guides me. The solutions worth building are the ones you can measure, correct and sustain over time.
🇦🇷 Español
Soy investigador independiente en inteligencia artificial y hace más de 20 años que resuelvo problemas donde se cruzan el hardware, el software y el mundo físico. Investigo cómo lograr que un modelo de lenguaje recuerde lo que se le dijo, lo corrija cuando cambia y diga «no sé» ante lo que nunca escuchó, con experimentos preregistrados y resultados publicados, también los que fallan. De día dirijo Westo, donde llevo esa misma forma de trabajar a obras e instalaciones industriales, y construyo mis propios sistemas de trading cuantitativo. Me guía una idea simple. Las soluciones que valen la pena son las que se pueden medir, corregir y sostener en el tiempo.
Independent research on memory, capacity and convergence in sequence models. All records are open access.
| Year | Work | DOI |
|---|---|---|
| 2026 | Stopping Criteria Below the Signal-to-Noise Floor: Window Length, Not Tolerance, Governs Convergence Detection in Architecture Comparisons | 10.5281/zenodo.21630279 |
| 2026 | stoppower — size your early-stopping window by statistical power (software · PyPI) | 10.5281/zenodo.21711767 |
| 2026 | «Ligamento» — A pre-registered, frozen experimental protocol (v1.0) for specialization vs. sharing in small-scale transformers | 10.5281/zenodo.21495252 |
| 2026 | The surprise stream warns before fast-weight memory collapses — and its autocorrelation does not (code & data) | 10.5281/zenodo.21385806 |
| 2026 | Cortex-Nexus: Domain-Specific Emotional Prompt Engineering for Large Language Models | 10.5281/zenodo.19866195 |
- 🔬 GENESIS — Private experimental research on physics-consistent modeling for complex material systems.
- AI-assisted modeling of physical systems
- High-density / long-retention memory concepts
- DevNet automation & systems integration
| Repository | Description |
|---|---|
| 🧬 telar-ligamento | Pre-registered protocol and results on specialization vs. sharing in attention heads. |
| 🧠 ReactionNet | Siamese neural network for chemical reactivity prediction. |
| 🧪 Periodic Table Dataset | Structured dataset for ML-driven materials research. |
| ⚙️ DevNet2025 | Automation & network programmability labs. |
| 🚀 Sample-App | CI/CD pipeline demonstration with Jenkins and Python. |
| 📂 IT-Lab-Archive | Archive for early IT experiments and foundational systems research. |
All repositories are experimental unless stated otherwise. Content is structured to emphasize clarity, intent, and technical direction.
Python Cisco DevNet Network Automation Computer Vision YOLOv11 Deep Learning PyTorch TensorFlow FastAPI Docker Kubernetes CI/CD Jenkins Linux Bash System Architecture Physics Simulation Materials Science R&D
