FORM 3 · SARAWAK, MALAYSIA
I build systems that have to show their evidence.
I explore AI systems, secure software, data, digital hardware, and electronics. I turn questions into inspectable prototypes, then separate what is simulated, implemented, and measured.
| Project | Area | What it demonstrates |
|---|---|---|
| RecurQuant | AI inference research | Physically packed recurrent-state quantization for Qwen3.5. On a frozen 500-task teacher-forced MBPP confirmation, its mixed INT8/INT4 policy used exactly 2,564,096 resident bytes and reduced macro excess NLL by 72.75% versus uniform INT4. |
| SFTGuard | Fine-tuning reliability | Fail-closed dataset, mask-evidence, and paired regression gates. Its sealed synthetic suite detected all 270/270 required fault signals with 0/30 clean-control false positives. |
| StrataMoE Lab | AI systems research | Deterministic GPU–RAM–NVMe placement harness with provenance-bearing traces, including a pinned Switch-Base-8 capture. Its first captured-trace benchmark records the preregistered policy missing its traffic gate. |
| CyberRAG | AI + cybersecurity | Local threat-intelligence retrieval with hybrid search, ATT&CK grounding, citations, and a fixed 15-question paired evaluation harness. |
| Edge AI RTL Lab | Digital design | Signed INT8 SystemVerilog compute core checked across 369 deterministic transactions against a bit-exact Python model and Yosys structural synthesis. |
| Local Evidence MCP | Secure tooling | Constrained evidence server with five narrow tools, zero execution capabilities, allowlisted access, redaction, safe writes, and 18 executable checks. |
| DataTrust Gate | Data + full-stack software | Browser-based dataset release auditor with a 15/15 fixed detector regression spanning privacy signals, leakage, duplicates, labels, provenance, and licensing. |
- ScamShield AI — documented case study of a privacy-first Malaysian scam-risk prototype; its dated verification snapshot records 202 passing tests and a clean analyzer without presenting those checks as detection accuracy.
- CustodianMesh AI — full-stack federated decision-support simulation with a 30-case fixed policy regression and no unexpected capability calls.
- Shark Habitat Prototype — reproducible Streamlit exploration of environmental scoring with explicit separation between software behaviour and ecological evidence.
define the question → scope the claim → build → test → document the boundary
- AI outputs should identify the evidence they used.
- Software should expose security boundaries and failure paths.
- Digital hardware should agree with a reproducible software model.
- Simulated, implemented, and measured results should never be mixed together.
These are student-built prototypes and small evaluations—not production deployments, agency systems, or silicon results.
Calendar note: the LABEEB + Espeon pattern is intentionally generated contribution art, not a record of development activity.
Python · PyTorch · LLM quantization · TypeScript · Dart/Flutter · FastAPI · browser APIs · RAG evaluation · MCP · SystemVerilog · Yosys · GitHub Actions
I welcome technical feedback, mentorship, job shadowing, student programmes, and small supervised project opportunities.