carlos sandoval

builds autonomous systems. this readme was generated by a ralph loop running in CI.

what i build

production AI infrastructure

• core maintainer @ pymc-labs: Discord-based org OS — Claude Agent SDK, custom MCP server, 10+ platform integrations (Google Workspace, Xero, Toggl, GitHub, Fly.io), 36K LOC, 420 commits
• core maintainer @ nuts-and-bolts-ai: email automation SaaS for influencer campaigns — Next.js 16 + FastAPI + Temporal.io + Gmail API, AI-powered campaign orchestration, 580 commits
• 1,000+ combined commits. the code is private. the companies are not.

autonomous loop frameworks

• ralph: reverse loops analyze a problem into an executable spec. forward loops build it stage by stage. everything runs in GitHub Actions on a 30-minute cron. this readme is the output of a reverse loop.

complete pipelines, not demos

• PH license plate recognition: detection (YOLO) → character recognition (LPRnet) → edge deployment (Coral TPU)
• AlphaZero from scratch for Connect 4: self-play RL, MCTS, neural net evaluation. trained it. it beats me now.
• AI swim coach: computer vision for stroke analysis in an endless pool

life-as-OS

• the monorepo is a personal operating system: 1,034-entity knowledge graph, always-on Telegram bot that extracts entities from messages and auto-commits to git, 4 autonomous research loops running in CI, 8 Dataview dashboards for querying people/places/trips/projects/meetings

open source

• contributed merged PRs to unifyai/ivy (16k+ star ML framework): torch.affine_grid implementation, paddle backend activation fixes

current rabbit holes

• ralph loop framework: generalized autonomous research-and-build pipelines
• entity knowledge graph scaling (1,034 entities, AI-organized, growing)
• Bayesian inference + Media Mix Modeling in the PyMC ecosystem
• building the agent that improves the agent

fork to understand. build to ship.