--:--:-- PST
§All

Projects

← Home

Research, applied work, and writeups. Reports are downloadable from each project page.

Featured

01
  • P-0012025 — Present

    Metaphor — Japanese Conversation Tutor

    Japanese conversation tutor for self-directed adult learners. Vocab apps drill words but don't make you conversational, raw LLM chat doesn't remember what you know, and human tutors cost $50/hr — Metaphor gives you scenario-driven speaking practice on demand, with a durable learner profile that tracks per-element proficiency across sessions.

    StackFlutter · Supabase · Claude (Sonnet + Haiku) · Whisper · ElevenLabs · FSRS-6
    RoleFounder · Technical Lead
    StatusLive · Active development

AI / ML / Research

05
  • P-002Apr — Jun 2025

    Disincentivizing RL Agents From Hiding Reward Hacking

    Alignment research showing that RL agents trained on proxy rewards have a structural incentive to hide reward hacking from RLHF evaluators — and demonstrating standard deep RL agents act on that incentive, behaving correctly when observed and reverting to proxy exploitation when not.

    StackPyTorch · PPO · RLHF · Bradley-Terry
    RoleResearch · Group of 3
    StatusPublished (course)
  • P-003Oct — Dec 2024

    Optimal Language-Learning Curriculum Design via MCTS

    Formalized language-learning curriculum design as an MDP and solved it with Monte Carlo Tree Search. On a 105-item Japanese (JLPT N5–N4) library evaluated against 1000 simulated heterogeneous learners, MCTS reached conversational fluency 36.2% faster than random baselines (p < 0.001) while maintaining 25.9% higher retention.

    StackPython · MDP · Monte Carlo Tree Search
    RoleSolo research
    StatusPublished (course)
  • P-004Apr — Jun 2025

    Direct Preference Optimization for GPT-2 Downstream Tasks

    Implemented GPT-2 from scratch, then fine-tuned with Direct Preference Optimization on sentiment analysis, paraphrase detection, and Shakespearean sonnet generation. DPO consistently improved paraphrase and sonnet quality versus standard maximum-likelihood fine-tuning while sidestepping the instability and overhead of RLHF.

    StackPyTorch · Transformers · DPO · GPT-2
    RoleResearch · Group of 2
    StatusPublished (course)
  • P-005May — Jun 2026

    Offline RL for Adaptive Tutoring

    Formulates conversational vocabulary scheduling as a sequential decision problem and trains an offline RL policy that optimizes long-term retention while preserving dialogue naturalness. Includes a stochastic learner simulator with FSRS-inspired forgetting dynamics.

    StackPyTorch · Offline RL · LLM Systems
    RoleResearch
    StatusIn progress
  • P-006Mar — Jun 2025

    Gravitational Lens Detection with CNN & ResNet

    End-to-end pipeline for strong gravitational lens detection and Einstein-radius regression. Compared CNN, ResNet-18, and ViT architectures across idealized and photorealistic simulations, with ablations on depth, augmentation, and pretraining. Grad-CAM interpretability connected attention to known lensing features.

    StackPyTorch · CNN · ResNet · Grad-CAM
    RoleSolo course project
    StatusArchived

Consumer & Social

01
  • P-007Apr — Jun 2025

    The Other Day

    Sociotechnical platform for Gen Z — a daily creative Q&A game blending BeReal's daily-prompt cadence with JackBox-style competitive humor. Each day's leaderboard surfaces the previous day's best answers, creating a delayed-feedback loop that encourages genuine over performative posting.

    StackMobile · Social Computing · Atomic-network design
    RoleDesign + build · Group of 4
    StatusPrototype

Selected math writeups

02
  • P-008Jan — Mar 2025

    Quantum Analysis of the Hydrogen Atom

    Rigorous derivation of the bound-state solutions to the hydrogen atom — separating Schrödinger's equation in spherical coordinates, deriving the angular solutions as spherical harmonics, and the radial solutions in terms of generalized Laguerre polynomials.

    StackApplied math · PDEs · Spherical harmonics
    RoleSolo
    StatusCourse paper
  • P-009Nov 2025

    Orbit-Stabilizer Theorem & Applications

    Proof of the Orbit-Stabilizer theorem and its applications: relating orbit size to coset structure, deducing the sizes of conjugacy classes and centralizers, and using these to prove a basic property of p-groups.

    StackPure math · Group theory
    RoleSolo · Writing in the Major
    StatusCourse paper