MSc Artificial Intelligence · NTU Singapore
I build AI systems that people can rely on.
I'm an MSc Artificial Intelligence student at Nanyang Technological University and a graduate researcher in Prof. Erik Cambria's SenticNet Lab. I build AI systems end to end and make sure they hold up on real data: a live web and mobile platform with an LLM assistant, LLM evaluation and agentic systems, multimodal deep learning for city-scale climate modelling at A*STAR, and efficient fine-tuning and inference.
Before my master's, I spent 2 years at NielsenIQ delivering research for global consumer brands and leading an 11-person team, where I learned that AI is only as useful as the decisions people can confidently make with it.
Experience
Publication
SenticLens: A Production Sentiment-Analysis Framework with Behavioral Auditing of LLM Affective Coherence
Simerjit Kaur Nath, Erik Cambria, Haibo Pen · SENTIRE Workshop at IEEE ICDM 2026 (ICDM is ranked CORE A*) · Accepted, to appear in the Proceedings of the 2026 IEEE International Conference on Data Mining Workshops (ICDMW)
Designed a paired benchmark: a 4,000-review corpus across 4 brands, matching real reviews with synthetic reviews of known sentiment. Across 24,000 audited records and 3 open-weight LLMs, models contradicted themselves ~6× as often on real reviews as on clean data (2.32% vs 0.37%).
Selected work
SenticLens pipeline: checking LLM answers without labels
MSc research at NTU: built SenticLens into a 6-stage pipeline that collects reviews from Reddit, the App Store and Google Play, scores them with LLMs, and checks every record before it is counted. Its label-free consistency check needs no ground truth: on 10,000 public benchmark reviews with star ratings, answers it flagged were wrong 9–25× as often as the rest.
Kenza Living: community platform, live on iOS, Android and web
Led the end-to-end build of a bilingual English/Arabic platform: React Native app, Node.js and PostgreSQL on AWS, and an LLM assistant that turns residents' messages into bookings and service tickets, with server-side guardrails for both security (authentication, role checks, rate limits) and correct bookings (server-sourced details, explicit confirmation).
Agentbox: open-source contributor
A desktop app that runs many Claude Code and Codex agents in parallel, each in its own git worktree. First contributor to bring Windows support: traced why agents could not start on Windows, then fixed how Claude Code and Codex launch, how tool servers and per-task folders work, and the in-app terminal.
RAG grounding audit
Open-source test of whether RAG chatbots really answer from their documents (3 LLMs, 500 questions). Reordering the same documents flipped up to 1 in 7 answers.
AdaptLink: AI seller intelligence for Shopee
My product idea and seller dashboard: turns sales, reviews and competitor prices into a health score and margin-safe next actions. Regional Finalist, team of 4.
CodeLah! Efficient code-to-language translation
LoRA fine-tuning of a small 350M-parameter model: 100% valid syntax, +40.8% over the baseline.
Hardware-aware U-Net for Stable Diffusion
24% less memory and 12.6% faster inference, with almost no loss in image quality.
TabM++: extending a tabular deep-learning baseline
Controlled extensions to the ICLR 2025 TabM-PLR model, tested on 5 TabReD datasets with 3 seeds: improved the matched baseline on all 5.
BDI multi-agent system
6 adaptive agents that plan with A* pathfinding, split territory into zones and message each other, on grids up to 80×80.
Education
Honours and leadership
Contact
Turning data into decisions, AI into advantage and ideas into innovation · Singapore