I study how AI systems fail when surface-level sentiment masks deeper user intent. My research sits at the intersection of NLP, intent modelling, and LLM reliability — three preprints published on TechRxiv (IEEE) and SSRN, seven deployed AI systems built from research to production.
I am a CS graduate specialising in NLP research, LLM reliability, and the design of AI systems that go beyond surface-level language understanding.
My research investigates a specific failure mode — when AI systems respond to how something is said rather than what it means. Three preprints that formalise this gap and propose intent modeling frameworks that operate beyond sentiment detection.
Alongside the research, I build production systems that operationalise these findings — clinical NLP pipelines, hallucination benchmarking platforms, RAG systems with claim-level verification, and interview-coaching and journaling platforms driven by intent-modeling pipelines.
Most NLP systems are built to detect sentiment. What they cannot do reliably is infer intent — the difference between someone saying "I'm fine" and meaning it, and saying "I'm fine" while meaning something else entirely. My research formalises this gap, identifies its failure modes, and proposes frameworks that treat intent as a first-class signal rather than a downstream assumption from sentiment.
Studying failure modes in sentiment-aware systems, benchmarking LLM hallucination across model families, and building production-grade NLP pipelines grounded in published research frameworks.
Deployed implementation of my TechRxiv (IEEE) framework — a five-layer pipeline operationalizing sentiment detection, pragmatic classification, temporal trend recognition, goal alignment, and utility-based response selection in a live journaling platform.
Browser-based platform benchmarking LLM reliability on medical question answering, classifying hallucination types, and comparing prompting strategies across model families. Chain-of-thought prompting roughly halved the hallucination rate versus zero-shot on this benchmark.
Full RAG pipeline — hybrid dense (pgvector) + BM25 retrieval fused via reciprocal rank fusion, cross-encoder reranking, streamed grounded generation, and claim-level hallucination detection — with the retrieval layer exposed: chunk similarity scores, source attribution, and per-claim grounding status.
Interview practice and simulation platform where every spoken or typed answer is passed through a four-layer Multi-Layer Intent Model (MLIM) — affective sentiment, pragmatic speech-act classification, goal-state tracking, and intent fusion — alongside conventional question generation, scoring, and a browser-side integrity-monitoring layer.
4 of 7 deployed systems shown above — explore the rest on GitHub.
View All Projects on GitHub ↗️This paper analyses the limitations of sentiment-based AI systems and proposes a multi-layer framework integrating pragmatic reasoning, temporal pattern recognition, and goal-aware intent interpretation for reflective writing contexts.
Read PreprintThis research formally distinguishes sentiment detection from contextual intent interpretation in human-AI communication systems, proposing a richer framework for understanding user intent beyond surface-level sentiment signals.
Read PreprintThis research presents a dual-model sentiment analysis system combining LSTM neural networks with VADER lexicon-based analysis to compare creator transcript sentiment against audience comment sentiment across YouTube content domains, revealing systematic divergence patterns and documenting domain-specific failure modes.
Read PreprintOpen to research collaborations and AI/NLP projects. Feel free to reach out.