APPLIED AI ENGINEER · HONG KONG / AVAILABLE GLOBALLY
Applied AI that survives contact with real data — document intelligence, agents, evaluation, and edge deployment.
| δ(q, σ) | → (q', σ', M) | FAILURE MODE |
|---|---|---|
| δ(q0, 01) | (q1, 01, R) | DATA DRIFT |
| δ(q1, 02) | (q2, 02, R) | EMBEDDING VERSION DRIFT |
| δ(q2, 03) | (q3, 03, R) | SEMANTIC MISMATCH |
| δ(q3, 04) | (q4, 04, R) | CONTEXT-BOUNDARY DEGRADATION |
| δ(q4, 05) | HALT | COST-DRIVEN PERFORMANCE COLLAPSE |
Source documents, queries, and user inputs shift distribution from what the pipeline was tuned on. New formats, languages, and edge cases arrive faster than chunking and parsing adapt, so provenance and access-control metadata silently go missing.
// Each stage has a characteristic failure mode. Evaluation catches them before deployment.
Systems that survived user studies, evaluation gates, and production constraints — designed under ambiguity and shipped anyway.
A framework for authorial control and game-state consistency in LLM-driven interactive digital narratives — published at ACM CHI 2026 as my M.Phil. thesis work.
An LLM-driven authoring tool that balances authorial control with generative improvisation. Validated through user studies with 28 participants against Twine and AI Dungeon.
A 24-crate Rust neural network framework for auditable, backend-agnostic NLP research. No hidden global state, backend-independent model definitions, and reproducibility by construction.
A Windows Service that backs up SQL Server databases into a single vendor-owned Google Drive account. A gated validation pipeline proves restorability end-to-end before every upload.
A 24/7 self-improving agentic AI trader on Deriv and cTrader. A deterministic engine of ~30 indicators gates every trade; an LLM council sits around it as a tie-breaker and risk-only veto.
Benchmarks, regression checks, and failure criteria — so teams compare models on evidence, not intuition.
OCR, extraction, and retrieval workflows for messy documents — secure, controllable, and auditable from day one.
Papers to products: interfaces, demos, and internal tools your stakeholders will actually use.