🧬 TADIYOS ASCHALEW

ታዲዮስ አስቻለው

Software & AI Engineer · Ethco Coder · Founder

---

"𝐼 𝑑𝑜𝑛'𝑡 𝑏𝑢𝑖𝑙𝑑 𝑜𝑛 𝑤ℎ𝑎𝑡'𝑠 𝑎𝑙𝑟𝑒𝑎𝑑𝑦 𝑎𝑠𝑠𝑢𝑚𝑒𝑑. 𝐼 𝑟𝑒𝑡𝑢𝑟𝑛 𝑡𝑜 𝑤ℎ𝑒𝑟𝑒 𝑡ℎ𝑒 𝑎𝑠𝑠𝑢𝑚𝑝𝑡𝑖𝑜𝑛𝑠 𝑏𝑒𝑔𝑎𝑛—𝑎𝑛𝑑 𝑟𝑒𝑏𝑢𝑖𝑙𝑑 𝑓𝑟𝑜𝑚 𝑡ℎ𝑒𝑟𝑒."

I'm a 20 y/o self-taught engineer from Ethiopia building the next generation of intelligent systems from first principles. No institution handed me this path. What drives every line of code here is a single conviction: the foundations most people accept as fixed can, in fact, be rebuilt.

Current AI hits hard ceilings—transformers scale quadratically, models forget what they learn, systems correlate but don't reason about causation. These aren't engineering problems. They're architectural problems requiring rethinking from the ground up.

That's what I do. I find the broken foundation. Understand why it broke. Rebuild it properly.

• Purpose: A fifth-generation programming language built around persistent, stateful computation.
• Core capabilities:
  • Orthogonal persistence keeps variables available across restarts, power failures, and reboots without explicit serialization.
  • Register-based virtual machine with NaN-boxing.
  • Neural-native vectors as first-class values.
  • Dimensional typing that enforces physical relationships at compile time.

var session_count = 0;  // Persists automatically
session_count = session_count + 1;
print "Session #" + session_count;  // Survives across restarts

• Purpose: A biologically plausible spiking-neural-network simulator designed for large cognitive models within a fixed memory budget.
• Scale: 270,336 neurons and 13,516,800 synapses within 500 MB of RAM.
• Architecture: Universal Intellectual Neuron (UIN) building blocks with six neuron classes and five synapse types.
• Performance: O(1) update kernel reported at 94× real-time speedup, with a meta-cognitive controller for adaptive resource allocation.

• Purpose: A hybrid C++/Python image-processing engine that converts visual data into mathematical logic atoms.
• Representation: RGB matrix, Energy (luminance), and Flow (gradient) layers.
• Performance: Sub-200 ms processing for 8K UHD images, reported throughput of 508 MB/s, and lossless reconstruction with PSNR above 66 dB.

• Purpose: An autonomous multi-agent orchestration engine for complex, end-to-end workflows.
• Specialists: Sixteen domain-focused agents covering language, analysis, innovation, frontend and backend development, debugging, QA planning, code quality, security, performance, UX, adversarial testing, and documentation.
• Coordination: Meta-cognitive planning connects specialist outputs into a collaborative development process.

• Purpose: A production-oriented framework for training decoder-only large language models from scratch.
• Model components: RoPE embeddings, RMSNorm, SwiGLU, and Multi-Query Attention.
• Training lifecycle: Pre-training, LoRA fine-tuning, and RLHF alignment through PPO/DPO.
• Language support: Integrates EthioBBPE for Amharic and Ge'ez tokenization.

• Purpose: A C++20 numerical kernel for metric-aware attention geometry.
• Operations: Builds learned positive-definite metrics, computes symmetric whitening operators, and transforms query/key coordinates.
• Design targets: Deterministic F32 CPU tensors and context lengths up to one million tokens.
• Validation: Forward-path checks include numerical guards and condition-number bounds.

• Purpose: An evidence-gated research implementation of selective cohomodynamic recurrent state-space models for causal language processing.
• Current implementation: CDI v3 uses the EthioBBPE tokenizer, sparse graph-Laplacian correction, and tied output projection.
• Reported results: CCT-G3.4 showed a 21% throughput improvement from token-residual evidence; CCT-G3.6 achieved bounded continuation at 6.576891 validation loss and outperformed GRU baselines across all seeds.
• Verification: 297 tests pass, with geometry re-confirmed through the G3.1–G3.6 gates.

I work where current approaches plateau. If you care about:

• 🎯 Rebuilding AI foundations instead of scaling broken assumptions
• ✓ Mathematical proofs over empirical approximations
• 🌍 Technology serving underserved language communities
• 🚀 Production systems that actually ship
• ⚛️ Quantum-classical computing on available hardware
• 🧠 Cognitive architectures that accumulate knowledge

Let's talk.

📧 nexuss0781@gmail.com

Self-taught. Community-rooted. Foundation-first.

Find the broken foundation. Understand why it broke. Rebuild it properly.

---

🌟 Connect

---

📊 GitHub Stats

---

🎨 Visual Collection

Profile Badges

AI/ML Stack

Research Status

Neural Architecture

Languages & Compilers

Agents & Automation

Security & Infrastructure

Applications

Dynamic Stats

Star History

---

Built from First Principles · Powered by First Principles

© 2024 Tadiyos Aschalew · All Foundations Rebuilt