MBZUAI AI Engineering undergraduate
I am focused on LLM systems, agentic workflows, and model evaluation, with experience building data pipelines, PyTorch prototypes, and reliability-oriented AI applications. My current interests include statistical learning, sequence modeling, decision systems, NLP, and reliable LLM deployment.
Selected work
LLM inference policy evaluation
First-author ICML 2026 Hypothesis Testing Workshop project designing statistical evaluation methods for LLM inference policies under latency deadlines, separating leaderboard gaps from certified policy-level advantages.
LLM hidden-state geometry
First-author ICML 2026 Mechanistic Interpretability Workshop project analyzing layerwise hidden-state geometry, showing raw L2 expansion is driven by norm growth while centered cosine better reveals semantic, topic, and category structure.
Natural-language scheduling product
Built an application that converts natural-language scheduling requests into Google Calendar and Outlook events, including preview, edit, delete, multi-event requests, recurring schedules, relative dates, timezone handling, undo/redo, command persistence, auth, and API validation.
Controlled LLM generation
Ongoing ICML 2027-oriented project developing a controlled-generation system combining RLM-style hidden-context access with Defect-Guided Verifier Search over visible candidates, with typed schemas for facts, forbidden claims, constraints, verifier gates, repair targets, candidate logs, branch budgets, and token accounting.
Repository link withheld while the project is being prepared for an ICML 2027 submission.
NLP and digital humanities
Ongoing ISMIR 2027-oriented project building a Songci NLP pipeline over classical poetry JSON corpora, extracting cipai, line length, rhyme position, ping/ze tone, and rhythmic features into PyTorch sequence-modeling inputs.
Repository link withheld while the project is being prepared for an ISMIR 2027 submission.
Symbolic reasoning verification
Co-developed step-level symbolic verification for LLM reasoning, using constraint propagation to check atomic deductions and strengthen constraint-dense logical reasoning.
Workshop paper at IJCAI/ECAI 2026 NILA.
Current focus
How should LLM inference policies be evaluated when latency, tokens, and compute are limited?
When do layerwise hidden-state distances support falsifiable semantic interpretations?
How can controlled generation improve validity without hiding quality or cost tradeoffs?
What can sequence models reveal about structure in classical Chinese poetry?
Signals