Hi, I’m Medhavi — a systems-focused AI builder crafting intelligence from first principles. I don’t just train models — I forge thought, shape logic, and challenge assumptions. Every project here is a blueprint toward reasoning-based AGI.
| Project | Description | Status |
|---|---|---|
| 🔧 CortanaPlusPlus | Custom C++/CUDA ML engine from scratch. No PyTorch. No TensorFlow. Pure control. | 🛠 In Progress |
| 🧠 RILITH | Reinforcement-Induced Logic in the Thinking Hierarchy — PPO-based AGI reasoning. | 🧪 Design Finalized |
| 🌱 Darwins-Silicate-Organism | Evolution-inspired simulation of artificial lifeforms learning memory, curiosity, and survival. | ⏳ Design Finalized |
| 🧬 TwinSqueeze | Siamese network with NEFTune noise for compressing MiniLM embeddings from 384 to ≤64 dims. | ✅ Complete |
| 🧩 GreedyContext | Compresses LLM input context by tracing the most relevant semantic path. | ✅ Complete |
| 🌐 Stream-JSON | Real-time parser for partial or token-level JSON/YAML LLM outputs. | ✅ Complete |
I believe AGI won’t emerge from brute-force training — but from ecosystems of AI modules that learn to reason, remember, and survive. My work combines:
- 🧠 Step-wise RL instead of token prediction
- 💡 Modular memory, curiosity, and reward models
- 🔁 Ecosystems of small interacting AI modules — not just one monolithic net
These projects are part of that roadmap.
- 🎥 A video showing how I train RILITH using a custom logic-based PPO setup
- 🧪 Dataset generation pipeline + reasoning model walkthrough
- 📓 Google Colab with annotated reward logic and RL charts
- 🧠 GitHub Projects
- ✉️ monishmedhavi@gmail.com
- 🌐 My Blog — reflections, architecture breakdowns & experimental logs
“Forge the mind, not just the model.”