McCoy R. Becker / Homepage / GitHub / MIT ProbComp

I work on probabilistic algorithms and software for modern hardware, with a lot of my time spent thinking about programming languages, compilers, and systems. I'm currently a PhD student at MIT ProbComp.

Current focus:

• probabilistic programming on accelerators, especially parallel ones
• programmable inference systems
• agent harnesses
• Human/agent interface and interaction design (an important area!)

I'd characterize myself as a systems thinker — I like sharp programmable tools. Most of my work starts from my desire for some sort of experience or useful item, and then I spend a lot of time thinking and tinkering, often with the goal of constructing one of these tools. I often get stuck in this phase, and like for my collaborators to yank me out. When it comes to projects, I'm not risk averse, and I'm willing to spend a large amount of time to get something right or try something crazy. This may indicate I'll never be a very successful academic, but I'm happy to share designs which I think are good with the world.

Local churn

Packages with recent commit activity.

glom agent context search

codebase analysis

sdfii terminal 3D SDF renderer

pelican constraint-based diagrams

Selected projects

Probabilistic programming

• genjax — Probabilistic programming language built on JAX, centered on generative functions and traces; supports programmable Monte Carlo + variational inference workflows, and is packaged as a POPL'26 artifact with docs/tests/case studies.
• programmable-vi-pldi-2024 — Artifact repository for the PLDI 2024 paper Probabilistic programming with programmable variational inference; includes the JAX implementation plus reproducibility scripts for the evaluation figures/tables.
• Jaynes.jl — Research-oriented alpha DSL/compiler for probabilistic programming in Julia; uses source-to-source IR transformations and contextual dispatch to implement the Gen.jl generative function interface.
• Problox.jl — Julia DSL for probabilistic logic programming that wraps ProbLog; compiles Julia-side model syntax into ProbLog programs and supports direct evaluation/querying from Julia.

Language and compiler systems

• abstraps — Rust compiler framework for extensible abstract interpretation, with MLIR-isomorphic IR concepts, abstract-VM interpreter interfaces, and builders for MLIR code generation.
• juju — Extensible compiler from JAX computations (Jaxprs) to Modular MAX graphs; lowers JAX primitives to MAX operations for execution in MAX inference sessions.
• pistachio — Experimental dependently typed language/theorem-prover implementation focused on elaboration and normalization-by-evaluation techniques.

Agent tooling

• mu — Programmable personal assistant for technical work, designed for long-running execution/persistence/reactivity; exposes shell-first primitives (issues, heartbeats, programmable mu_ui docs) for custom orchestration.

Papers

• Probabilistic programming with vectorized programmable inference (POPL 2026)
• Probabilistic programming with programmable variational inference (PLDI 2024)
• Reverse-mode ADEV via YOLO: tangent estimators transpose to gradient estimators (LAFI 2024)
• Brain-wide representations of behavior spanning multiple timescales and states in C. elegans (Cell 2023)

Tools I use

Open-source software from my dotfiles that makes my workflow possible. Nodes with a ♥ link to the project's sponsor page.

_{click to explore — nodes with ♥ link to sponsor pages}

---

femtomc.github.io