AVA · Pantheon-Trades · MALD · pane · MAVIS · Terminus
Building local-first AI systems — fine-tuning, evaluation, multi-agent orchestration — and the knowledge tools and shell environments around them. Every claim below links to running code. BSc (Hons) IT at Middlesex University Dubai.
📌 Open to AI / ML / Software Engineering internships and junior roles — Dubai / Abu Dhabi or remote · UAE Golden Visa (no sponsorship needed) · immediate start · m.afsah.279@gmail.com
flowchart LR
A["Project Portfolio"] --> E["AVA — Local AI Lab<br/><i>Python · PyTorch · QLoRA · 4 GB VRAM</i>"]
A --> P["Pantheon-Trades — 11-Agent Trading Council<br/><i>Python · FastAPI · Solidity · Next.js · LIVE</i>"]
A --> B["MALD — Markdown Workspace<br/><i>Rust · GUI/TUI/CLI · Local AI + cited RAG</i>"]
A --> N["pane — Linux Desktops on Windows<br/><i>Rust · WSL2 · XRDP</i>"]
A --> D["MAVIS — Windows Shell Environment<br/><i>Rust · Dear ImGui · Lua</i>"]
A --> C["Terminus — Terminal Desktop<br/><i>C++ · wxWidgets · Lua</i>"]
style A fill:#161b22,stroke:#8b949e,color:#c9d1d9
style E fill:#0d1117,stroke:#f78166,color:#c9d1d9
style P fill:#0d1117,stroke:#3fb950,color:#c9d1d9
style B fill:#0d1117,stroke:#58a6ff,color:#c9d1d9
style N fill:#0d1117,stroke:#58a6ff,color:#c9d1d9
style D fill:#0d1117,stroke:#8b949e,color:#c9d1d9
style C fill:#0d1117,stroke:#8b949e,color:#c9d1d9
| Layer | Project | Core Tech | Receipts |
|---|---|---|---|
| AI Research | AVA | Python, PyTorch, PEFT/QLoRA, GGUF | Fine-tuned a 2B LLM to 82.0% ARC-Challenge / 92.0% ARC-Easy — beating Llama 3.2 3B-Instruct (78.6%) — with a 42 MB adapter on a single 4 GB GPU. 17-benchmark, 16,872-task eval harness. |
| Multi-Agent | Pantheon-Trades | Python, FastAPI, Solidity, Next.js | Live: pantheon-trades.vercel.app — 11 LLM agents debate every prediction-market trade. Half-Kelly sizing, Platt/isotonic calibration, Brier 0.149 on a 200-market backtest, 714 Python + 65 Solidity tests. |
| Knowledge | MALD | Rust, SQLite FTS5, HNSW, Ollama | Single-binary PKM: hybrid full-text + semantic search, cited RAG chat down to exact file and line. Fully local. |
| Environments | pane | Rust, WSL2, XRDP, PowerShell | Real Linux desktops on Windows — create, launch, repair from one control surface. Arch-first MVP shipped as pane.exe. |
| Interface | MAVIS | Rust, Dear ImGui, Lua | Scriptable Windows shell replacement: terminal, editor, widgets, system monitor in one environment. |
| Interface | Terminus | C++, wxWidgets, Lua | Windows terminal environment replacing the desktop with widgets, search, and file workflows. |
| Domain | Skills |
|---|---|
| AI / ML Engineering | LoRA/QLoRA fine-tuning, model evaluation & benchmarking, RAG pipelines, multi-agent orchestration, quantized inference (GGUF), model serving — PyTorch, Hugging Face, Ollama |
| Backend & Data | FastAPI, Node.js, REST API design, PostgreSQL, Redis, SQLite FTS5, vector search (HNSW) |
| Blockchain | Solidity, Foundry, on-chain settlement (Arc Testnet), Polymarket CLOB integration |
| Native Systems & UI | Rust and C++ application architecture, Windows shell design, terminal-first UX |
| Web & Real-Time | Next.js, React, Three.js, WebGL, Socket.IO, shader-driven visual systems |
| Tooling & Delivery | Docker, GitHub Actions CI/CD, CMake and Cargo workflows, cross-platform scripting, release packaging |
- 🔬 Open-source research contributor — HAGI: Clifford-algebra reasoning in small language models (PyTorch, Rust, Lean4)
- 🏆 2× hackathon runner-up — Odoo Hackathon Dubai · Curtin University Dubai (Nov 2025)
- 🪑 Chairman, MDX Computing Society · Committee Chair, BCS Student Chapter
Making 2B models punch above their weight · keeping 11 agents arguing productively about markets
m.afsah.279@gmail.com · LinkedIn · Portfolio · Pantheon Live
