02 / 04

Work

Companies and systems I've helped build

Now building

Company

2025– · Cofounder & CEO · San Francisco

Virtual cellsBiologyCompany

Problem: Biology teams can map what exists and still struggle to decide which biological change is worth pursuing next.
Approach: Use virtual cell models to compare possible interventions before expensive experiments begin, while keeping computational claims answerable to biology.
What shipped: As cofounder and CEO, I am building Galen: a computational layer for understanding cellular behavior and focusing experimental work.
Evidence: The public product and First Principles publication make the current thesis, methods, and direction inspectable.

Shipped

2026 · MLX · autonomous research system

Autonomous loopsMLXGit-auditedLocal compute

Problem: Research iteration still consumes sustained human attention: propose a change, run it, inspect the result, and decide what survives.
Approach: Turn that loop into an auditable system with a fixed budget, one honest metric, isolated artifacts, and Git-backed keep-or-revert decisions.
What shipped: Bellwether is an MLX port and extension of Karpathy's autoresearch pattern for proposing, running, scoring, and auditing experiments on Apple Silicon.
Evidence: Public runs improved fixed-budget language-model training from 2.667 to 1.807 bpb in the initial walk; longer M4 Max sweeps reached about 1.29 bpb.

Shipped

2026 · Rust · open-source bioinformatics

HN front page185 pointsPublic verifierRust

Problem: Genomics pipelines often assume large machines, cloud workflows, and memory ceilings that are difficult to inspect.
Approach: Treat memory as a contract: declare the budget, stream the work, and produce a receipt that others can verify.
What shipped: Rosalind is a Rust genomics engine and verifier for whole-genome workloads on ordinary laptops.
Evidence: Someone else submitted the project to Hacker News, where it reached the front page at 185 points; the verifier, repository, and crate are public.

Shipped

2026 · Rust · exact dynamic-programming traceback

Exact tracebackBiosequence alignmentVerificationRust

Problem: Classic dynamic-programming traceback stores an entire table, making exact alignment expensive as sequences or frontiers grow.
Approach: Summarize and compose intervals like the original recurrence so the exact path can be reconstructed without retaining every cell.
What shipped: A Rust engine and hcp-align CLI for exact biosequence alignment, traceback, batch processing, and verification.
Evidence: The public alpha supports edit distance, Needleman-Wunsch, Gotoh, Smith-Waterman, semi-global alignment, FASTA/FASTQ batches, and optional SHA-256 result certificates.

Shipped

2026 · Python · Git for long stochastic runs

Bit-exact replayO(√T) memoryPython

Problem: Long stochastic simulations are difficult to inspect after the fact without retaining the entire trace.
Approach: Make a seeded run reproducible and branchable like Git history, even when the full history is too large to keep.
What shipped: A Python engine for recording, replaying, and branching multi-million-step stochastic runs in O(√T) memory.
Evidence: The public repository demonstrates bit-exact replay and traceback without materializing the complete run.

Shipped

2025 · Rust · sublinear-space ZK

Sublinear spaceZero knowledgeRust

Problem: Proof generation is usually memory hungry, making long proofs difficult on fixed-memory hardware.
Approach: Structure the prover as a streaming computation so commitments and openings can be assembled without holding whole polynomials.
What shipped: A Rust sublinear-space zero-knowledge prover with a public implementation and accompanying paper.
Evidence: The repository exposes the implementation and demonstrates that production-style commitments and sublinear space can coexist.