(2026) clue-llm
(live demo) -
Stress-testing LLM reasoning through next-token prediction. Generates
NP-complete SAT logic puzzles rendered detective-style as “who’s the
killer?” games. Uses SymPy’s SAT solver and asks LLMs solve them. Uses
single-token strings for clean confidence measurement, to see if models
“know when they’re right”/“wrong”. Includes experiments with strategic
fine-tuning on confident mistakes. Inspired by Ilya Sutskever’s claim
that predicting the killer in a detective novel requires “a fair amount
of reasoning.”
(2026) group_there
- Carpool optimization app for teams. NextJS/TypeScript
re-implementation of the original GroupThere carpool optimization app
(see 2016 entry below), with a Python solver backend deployed on Modal infrastructure. Plan is to
re-implement with cuOpt
to run on GPUs.
(2025) cuopt-stubs
- Python stubs for cuOpt LP/MILP solver. The full cuOpt can only be
installed on machines that support Cuda, which excludes MacOS - these
stubs support writing cuOpt programs with full type-checking support on
MacOS. Stubs auto-generated via MyPy stubgen.
(2025) fzf-ts - A
typescript interface to fzf via stdout and temp files.
Because every selection CLI should really just be fzf.
(2025) linear_cli
- CLI tool for managing Linear issues from the command line. Uses
fzf-ts. Implementation in Go: linear-cli-go.
(2025) taste-lever
- Simple one-pass prompt-optimization from human-annotated data.
(2024) FastRPC a
thin abstraction on top of FastAPI that, along with openapi-typescript-sdk-generator,
allows for a fully type-safe Typescript SDK to be generated instantly on
every save during development. Call Python functions from a fully
type-aware TypeScript client with no networking code.
(2023) Capsule
(live app on Vercel) -
a highly-stylized message sharing system. 100% type-aware full-stack
NextJS application (via server actions).
