Knowable
Most learning apps give everyone the same path. Knowable builds a short course around one person’s goal, then mixes explanations with predictions, labs, and conceptual checks.
Gemini 2.5 Flash · React · Cloudflare Workers
Source ↗Hi — I’m Sawyer.
I build software for questions I can’t leave alone.
Right now that means personalized learning, agents that actually do things, molecular machine learning, and ways to understand aging before it turns into disease.
01
Selected projects
Things I built because I wanted them to exist.
CrushIt
A goal agent that does not spend all day talking about your goals. It picks one concrete action each morning, checks whether you did it, and can attach a simulated commitment penalty.
Node.js · Maritime · OpenClaw
Source ↗WildAI
A wilderness assistant for the moment you have no signal. Point your phone at gear or terrain and a small local vision-language model tries to figure out what is around you and what to do next.
React Native · llama.cpp · on-device VLMs
Private prototypeLongevitree
A wellness prototype for turning a few simple at-home tests — reaction time, walking pace, sit-to-stand — into a rough picture of functional-aging signals and what might improve them.
Python · FastAPI · functional aging
Source ↗02
Research
Work at the boundary between models and experiments.
Can neural potentials trained on massive computed datasets predict real redox measurements?
We tested OMol25-trained neural network potentials on experimental reduction-potential and electron-affinity data, then compared them with cheaper quantum-chemical baselines.
Benchmarking OMol25-Trained Models on Experimental Reduction-Potential and Electron-Affinity Data
Read preprint ↗How far can a simple, interpretable rule go on a messy chemistry problem?
Using about 13,000 single-stereocenter compounds with experimental optical-rotation signs, we use machine learning to optimize and test a simple substituent-ranking rule.
Machine learning optimization of a simple rule for predicting the sign of optical rotation of chiral compounds
Read abstract ↗Could a virus be engineered to preferentially find and kill senescent cells?
A proposal combining uPAR-targeted infection, GLB1-controlled expression, and siRNA knockdown of candidate senescent-cell anti-apoptotic genes.
Selective viral targeting and clearance of senescent cells
The proposal has three aims: test uPAR as an infection target, test the GLB1 promoter for senescence-selective expression, and rank 18 SCAP-associated genes by how strongly their knockdown induces apoptosis in senescent versus non-senescent cells.
03
About
I like projects where the hard part is not making another interface — it is figuring out something real about molecules, cells, people, or how software can actually change behavior. I usually learn fastest by building the thing I wish already existed.