projects

Shipped · Open source · Python CLI & library · PyPI

ProofRAG

A reproducible evaluation loop for RAG systems: generate golden sets, score retrieval and generation, and block regressions in CI.

  • Python
  • CLI
  • GitHub Actions
  • PyPI
  • 188 tests

Shipped · Open source · Standalone Python scripts

rag-utils

Reusable pieces from my RAG pipeline work: hybrid retrieval, reciprocal-rank fusion, query expansion, embedding caching, and ingestion, evaluation and tracing utilities.

  • Python
  • Retrieval
  • Standalone scripts

Shipped · Android · On Google Play

Tether

A local-first reader for articles, feeds, PDFs and EPUBs. Read the whole page or one word at a time.

  • Kotlin
  • Jetpack Compose
  • Readium

Shipped · Real-time multiplayer · Deployed PWA · Co-built

Hukka Bukka

A bluffing party game built with friends: everyone writes a fake answer to the same question, then races to spot the real one. Fooling a friend scores almost as well as knowing the truth.

  • Next.js
  • Firebase
  • Zustand

Upstream Contributions · zed-industries/zed · Merged PR · Rust

Zed: JSONL/NDJSON previews

Merged PR to the Zed editor adding JSONL/NDJSON previews, edit recovery and a CSV crash fix, with docs and regression tests.

  • Rust
  • Zed
  • Parsing
  • Regression tests

Research · Top 10% · Digital Minds Research Sprint

SamplerScope

An exact paired audit of how decoding rules change the action policies induced by fixed language model logits.

  • Python
  • Qwen2.5
  • Finite MDPs

Research · Exploratory experiments wip

Prescriptive Bias in LLM Sampling

Do language models sample from statistical reality, or from their sense of how things ought to be?

  • pydantic-ai
  • Python
  • GPT-4o mini

Coursework & Earlier · Course project · Neural Networks

Data Selection and PEFT

How should a limited external molecular dataset be selected when the model itself is only partially fine-tuned?

  • MolFormer
  • PyTorch
  • RDKit

Coursework & Earlier · Course project · Generative AI

Program Repair and Hint Generation

Language models that repair introductory Python programs and write useful hints without giving the solution away.

  • GPT-4o-mini
  • Phi-3-mini
  • LoRA

Coursework & Earlier · Evidence-aware claim verification

Fact Check LIAR

Hybrid retrieval and classification that finds related political claims, predicts a six-level truth label, and explains it.

  • BM25
  • FAISS
  • BERT

Coursework & Earlier · Bachelor thesis

VoxArt

A speech-to-image system that turns spoken language, and the sentiment in it, into visual art.

  • BERT
  • Stable Diffusion
  • PyTorch

Coursework & Earlier · Weakly supervised computer vision

Scribble-Supervised Segmentation

Recovering dense object masks from a few foreground and background strokes with a four-channel U-Net.

  • PyTorch
  • U-Net
  • EfficientNet-B2

Coursework & Earlier · Cost regression · high-use classification

Healthcare Utilization Prediction

Two questions about the same patients: how much will care cost, and who will use a lot of it?

  • scikit-learn
  • XGBoost
  • Pandas