Research

Research projects

Papers, experiments, and system prototypes focused on AI reliability and learning.

Student writing notes in a notebook during a study session

2026 Feb - 2026 May · Research

10xLearners: Multi-Agent Personalized Learning

A multi-agent tutoring system that builds concept graphs, adaptive study plans, and assessment loops from learner-owned documents.

Technology: LangGraph, Qwen 2.5, Mistral API, Google Gemini API, Hugging Face Inference, Neo4j, MySQL, Google Drive API, Next.js, TypeScript, Tailwind CSS, FastAPI, Python, Alembic, JWT Auth

Algorithm: Knowledge Graph Construction, Graph-Aware Retrieval, Mastery-Based Adaptive Planning, Schema-Constrained Generation, Coverage-Driven Regeneration

Code editor on a laptop screen showing model analysis work

2026 Feb - 2026 May · Research

Residual-Controlled SAE Diffing for LLM Failure Discovery

A three-stage interpretability pipeline that improves LLM failure-pattern discovery by controlling for question difficulty, diffing SAE latents, and scoring the discovered pattern set with set-level metrics.

Technology: Sparse Autoencoder (SAE), Residual Analysis Pipeline, LLM API Inference, MMLU, Feature Artifacts, Evaluation Reports (HTML), Python, PyTorch, Transformers, Pandas

Algorithm: Residual Error Modeling, Residual-Based Failure Grouping, Sparse Autoencoder (SAE) Diffing, Latent Feature Attribution

Computer screen displaying code and model experimentation

Spring 2025 · Research

Learning When to Think: RL for Adaptive Reasoning Computation in LLMs

LLMs benefit from extra test-time computation, but fixed reasoning budgets waste compute on easy problems and still under-explore hard ones. We train a three-action policy (continue, refine, terminate) with GRPO, Adaptive Length Penalty (ALP), and DeGRPO-style control-token weighting on Qwen2.5-Math-7B-Instruct (LoRA), and evaluate on MATH-500 with accuracy, efficiency, and action-usage analysis.

Technology: Qwen2.5-Math-7B-Instruct, LoRA, GRPO Training Pipeline, MATH-500, Python, PyTorch, Transformers

Algorithm: GRPO, Adaptive Length Penalty (ALP), DeGRPO-Style Control Token Weighting, LoRA Fine-Tuning

Printed newspaper pages representing text categorization research

2012 · Research

News Categorization Using SVM

Undergraduate research (2012) on automatic Indonesian news categorization using TF-IDF features and Support Vector Machines.

Technology: TF-IDF Features, Labeled News Corpus, Java

Algorithm: TF-IDF, Support Vector Machine (SVM)