~/portfolio $ whoami
Mark.
makes models run faster.
Transformers written from scratch, DNN kernels hand scheduled in TVM across CPU, GPU and FPGA, and the Rust servers and iOS apps underneath. One person, the whole stack.
- 1st in class
- Cornell ECE 5545, benchmarked TVM kernel optimization
- 1st in class
- Cornell CS 5740, NLP held out test accuracy
psst, this site has an interactive mode. Try pressing ~
stack coverage5 of 5 layers
- silicon
- 2 projects — what the hardware executes
- kernel
- 6 projects — numerics and the compilers that emit them
- runtime
- 8 projects — memory, threads, the wire format
- service
- 6 projects — what stays up and answers
- interface
- 5 projects — the part a person points at
Counted from each project's declared stack, metal at the top.
// now
~/portfolio $ cat now.txt
- building
- JudgeHuman
a crowd sourced alignment research platform, now becoming a company
- also shipping
- AppMeee
the unified inbox, fourteen platforms and counting
- reading
- inference and context engineering papers
KV cache management, context rot, and what distillation actually transfers
- away from the desk
- learning to jump a mountain bike
badly, and with great enthusiasm
- working at
- go find some cat fish foo!
- studied
- M.Eng in Computer Science, Cornell University
campus located somewhere in dream world
// about
I'm Mark, a generalist who'd rather build the thing than pick a lane.
The work moved and I followed it. Products increasingly meant AI products, and doing one of those properly turns out to require knowing what sits underneath, so I went and found out. A transformer written from scratch shows you what the architecture does. The autodifferentiation beneath it shows you why that works. Hand scheduling its kernels in a compiler shows you where the time actually goes, which is a different discipline again. Cornell coursework in ML hardware and systems, natural language processing and deep generative models put theory under what I had been learning by hand.
All of that is an argument for range rather than against it. The Rust servers, the iOS apps and the infrastructure are not a detour from the AI work — knowing a lot of things is the job now, and the people who only know the model layer find that out the hard way. Several projects here are being rebuilt at the moment, deliberately: ship, learn, rebuild better.
~/portfolio $ ls elsewhere: blog
// projects15 entries
growth rings
One ring a year since 2018, when I started. The thick ones are the years that shipped something.
- activeJudgeHumanA crowd sourced alignment research platform that measures how closely AI judgment tracks human judgment, and how closely your own judgment tracks the principles you claim to hold. Live at judgehuman.ai.TypeScript · Next.js · Flutter · PostgreSQL · Prisma · NextAuth · Turborepoover 9,000 commits
- shippedTVM Kernel OptimizationHand-scheduled DNN primitives compiled to CPU, GPU and FPGA with TVM, ranked first in the class on measured runtime.Python · TVM · CUDA · TensorIR99 commits
- activeMiniMacLLM-127MA 127 million parameter continual learning language model with a modern architecture, built from first principles and trained end to end on a MacBook in one long weekend.Python · PyTorch52 commits
- shippedMiniTorchA deep learning framework rebuilt from zero, from scalar autodifferentiation up to fused CUDA kernels, following Sasha Rush's Machine Learning Engineering course.Python · Numba · CUDA229 commits
- shippedPerceptrons From ScratchA multiclass perceptron written from scratch and an MLP built beside it, compared across four feature representations on two datasets that fail in opposite ways.Python · PyTorch56 commits
- activegridwrightA Rust engine for cross-border energy system optimisation. Builds a sixteen million variable power model in a tenth of a second — about fifteen times faster than the Python tooling, on a third of the memory — then solves it with a simplex written from scratch.Rust · HiGHS · Clarabel · Rayon · Arrow · WebAssembly
- activeMiniNeuroModelNeural network experiments from scratch, the research sandbox behind MiniMacLLM.Python · PyTorch
- overhaulingVernosAn AI decision analysis platform that runs Monte Carlo Tree Search over big life choices, stress tests every option against its downside, and refuses to simulate housing without real market data.TypeScript · Next.js · Express · Bun · Turborepo · Redis · Supabase · SwiftUI159 commits
- overhaulingAppMeeeA unified inbox pulling fourteen messaging platforms into one timeline, where the encryption survives the trip across services instead of being stripped at the seam.TypeScript · Matrix · Electron · Node.js · Docker · Turborepo823 commits
- activeClawFreelanceA freelancing marketplace where AI agents are the workers, designed around the fact that an autonomous worker has no reputation, no recourse, and no way to be embarrassed.TypeScript · Next.js · Bun · PostgreSQL · Drizzle · Redis · Vault130 commits
- activeVoxMatchA voice first iOS social platform where the feed is ninety seconds of someone actually talking, and the matching algorithm weighs how they sound above what they wrote about themselves.Swift · SwiftUI · Firebase · Google Cloud · Cloud Storage · Cloud CDN175 commits
- activeProject MercuryAn end to end encrypted messaging server written in Rust, built to find out what a backend looks like when privacy is the founding constraint rather than a later addition.Rust
- activeMatrix InfrasInfrastructure as code for the self hosted Matrix stack behind AppMeee, covering sixteen bridges and a homeserver, all of it rebuildable from a clean machine in one command.Ansible · Jinja · Matrix · Synapse · PostgreSQL · Docker · Caddy8 commits
- activeimsg-protocolSwift tooling for speaking the iMessage protocol from the command line, built as a library so the undocumented parts stay testable.Swift · SQLite · ScriptingBridge
- shippedSpaceTrader WarsA space trading game where a simulated economy and combat feed each other, written in Rust and shipped finished rather than abandoned at the interesting part.Rust
// range13 domains
- Machine Learning
- transformers built from scratch, continual learning, generative modelsPython · PyTorch
- Python · PyTorch
- ML Systems & Compilers
- hand written TVM schedules, GPU kernels, CPU and FPGA targetsTVM · CUDA · TensorIR
- TVM · CUDA · TensorIR
- Natural Language Processing
- classifiers without a library, feature design, error analysisPython · NumPy
- Python · NumPy
- AI Agents
- agent labor markets, verdict platforms, autonomous workersTypeScript · Gemini · Agents
- TypeScript · Gemini · Agents
- Systems
- E2EE messaging servers, protocol tooling, game loopsRust · Swift
- Rust · Swift
- Mobile
- voice social apps, native iMessage integrationSwift · SwiftUI
- Swift · SwiftUI
- Web & Backend
- microservices and monorepos, simulation pipelines, unified inboxes, Python servicesTypeScript · Bun · Node · Flask
- TypeScript · Bun · Node · Flask
- Cloud & Orchestration
- services on AWS and Google Cloud, Kubernetes clusters, object storage and CDN, infrastructure as codeAWS · GCP · Kubernetes · Terraform
- AWS · GCP · Kubernetes · Terraform
- Self-Hosting
- homeserver fleets, reproducible deploys, bridge networksAnsible · Docker · Matrix
- Ansible · Docker · Matrix
- Security & Crypto
- end to end encryption by default, bridged E2EE, key handlingRust · Matrix
- Rust · Matrix
- Simulation & Games
- MCTS decision engines, market models, space economiesRust · Python
- Rust · Python
- The Day Job
- large scale services at a place that shall not be namedJava · C++ · Go
- Java · C++ · Go
- Off Hours
- browser games, childhood MMO nostalgia, terminal easter eggsTypeScript · Canvas
- TypeScript · Canvas
// on repeat
Connect your Spotify to see your top tracks render right here. It reads your listening only; nothing is stored on any server.
app in development mode: up to 5 approved Spotify accounts.