Form 3 student in Sarawak, Malaysia.
I build inspectable systems across AI inference and quantization, fine-tuning reliability, RAG, secure software, and RTL/digital design. I care about measured evidence, reproducible tests, and clear limits.
| Project | Focus | Evidence |
|---|---|---|
| CliffQuant (model) | Quantization and inference | Exact minimax FP16 scale selection for multi-environment W4A16 quantization. On frozen Qwen3.5-0.8B held-out windows, the validated checkpoint lowered macro NLL by 0.0548 versus uniform AbsMax. |
| RecurQuant | Recurrent-state quantization | A physically packed INT8/INT4 policy used 2,564,096 resident bytes and reduced macro excess NLL by 72.75% versus uniform INT4 on a frozen 500-task MBPP confirmation. |
| SFTGuard | Fine-tuning reliability | Fail-closed dataset, mask-evidence, and paired regression gates; the sealed synthetic suite found all 270 required fault signals with 0/30 clean-control false positives. |
| CyberRAG | RAG and evaluation | Local threat-intelligence retrieval with hybrid search, ATT&CK grounding, citations, and a fixed 15-question paired evaluation harness. |
| Edge AI RTL Lab | RTL and digital design | A signed INT8 SystemVerilog compute core checked across 369 deterministic transactions against a bit-exact Python model and Yosys synthesis. |
| StrataMoE Lab | AI systems research | A deterministic GPU/RAM/NVMe placement harness with provenance-bearing traces, including an honest captured benchmark where the preregistered policy missed its traffic gate. |
More builds: ScamShield AI, DataTrust Gate, Shark Habitat Prototype, Local Evidence MCP, and CustodianMesh AI.
scope the claim -> build -> test -> publish the evidence -> name the limits
These are student-built prototypes and small research evaluations, not production deployments or silicon results.
Working with: Python, PyTorch, LLM quantization, FastAPI, TypeScript, Dart/Flutter, RAG evaluation, MCP, SystemVerilog, Yosys, and GitHub Actions.
I welcome technical feedback, mentorship, job shadowing, student programmes, and small supervised collaborations.
The LABEEB + Espeon contribution pattern is intentional contribution art, not a record of development activity. Sprite source and attribution are documented in assets/README.md.