Projects

A selection of what I’ve built: products I’m taking to market, consulting work, open-source tools people actually use, and research systems. Most of it lives on GitHub.

Products

AWAM is the execution OS for Argentine fashion brands. An AI-first commerce platform that runs the full commercial loop: from a customer’s WhatsApp message to recommendation, checkout, payment, electronic invoice, stock deduction, and lost-demand recovery. Six specialized AI runtimes over a durable, multi-tenant core. Next.js · React · Prisma · Postgres · Mercado Pago · ARCA e-invoicing.

Quent is a real-time AI meeting copilot. Live transcription and a floating cognitive overlay that answers questions from your own documents (RAG) and generates memory cards while you talk. Electron front end over a multi-agent pipeline on Deepgram, Groq and Cerebras.

Reacher is an autonomous B2B prospection agent. It researches companies, drafts personalized outreach, sends, monitors replies, and follows up, unattended, at $0.06 per company. Custom TypeScript agent harness with observation masking, context compaction and cost tracking.

Consulting

sesgo.ai is my AI consultancy: machine learning, data science, AI agents, data engineering, and MLOps for teams across LatAm and the US. Recent work includes the Arison WMS Assistant, a conversational AI layer over a warehouse management system.

Open source & tools

linkedin-cli ★56. A terminal client for LinkedIn that works with real browser sessions instead of OAuth. Read feed, profiles and search; write posts, reactions and comments; rate-limit and proxy aware.

gralph ★16. A high-performance implementation of the Ralph autonomous loop. Turns a PRD into a task DAG and runs many coding agents in parallel, each isolated in its own git worktree, one PR per task.

trail. Shared, append-only memory for your team and its AI coding agents. Plain Markdown in your repo, opens as an Obsidian vault, links to your tracker. No DB, no daemon.

code-optimizer. A code-optimization audit that runs 14+ specialist agents in parallel, each hunting one class of performance anti-pattern via pattern matching, with auto-fix PRs.

autoagent. A system that iteratively discovers better agentic architectures, using TLA+ formal verification to check concurrent pipelines before spending tokens.

wa2vault. A read-only WhatsApp archiver that pulls chats into an Obsidian vault, transcribing voice notes locally with Whisper. Privacy-first and fully offline.

Research & ML systems

Vision foundation model for embryos (LiNAR). JEPA-style self-supervised representation learning over time-lapse IVF imaging, predicting embryo viability without a biopsy.

Hierarchical Reasoning Model. Work on a 27M-parameter, brain-inspired reasoning architecture with custom CUDA kernels that beats far larger models on ARC-AGI, Sudoku and mazes.

Training at scale. Training 70B-class LLMs on consumer GPUs with FSDP + QLoRA.

text2sql. Offline, CPU-only natural-language-to-SQL with MCTS refinement and grammar-constrained decoding.

This is a slice. Around 70 more repositories, from ARMv8 simulators and Unix-v6 filesystems to vision and NLP coursework, live at github.com/frizynn.