Private AI. Built around your work. Preparing quitereason.ai for businesses. Building personal AI agents for anyone, tailored to the way you work.
AI Engineering · ML · Web3 & Trading
github.com/ogarciarevettInteractive illustration · not live training
Simplified Transformer, bottom to top: embeddings, multi-head self-attention, feed-forward network and output projection. Residual connections, normalization and repeated blocks are omitted. Directional signals illustrate training; no actual LLM training or inference runs here.
Different needs. A shared focus: practical AI, built for real work.
Your knowledge. Your infrastructure. An AI built around your company.
I'm developing a self-hosted LLM service for businesses, with a focus on financial teams. Adapted with LoRA, it will pair your own model with an on-demand virtual Forward Deployed Engineer (FDE) for your team's workflows.
For anyone. A separate, bespoke service for your personal workflows.
Connect the tools you use. Turn recurring research, planning and repetitive work into an agent built around you.
Agents, tools and reliable execution.
Engineering the loop around the model: orchestration, retrieval, memory and evaluation.
Illustrative workflow — tools shown are alternatives, not one mandatory framework.
Plan → use tools → evaluate → iterate. Feedback changes the next action, not the model weights.
Compare overlapping agent skills in parallel, score outputs and combine the result.
CLI for updating cross-AI templates.
AI CLI templates with skills and pipelines.
Identity experiments for trusted personal AI agents.
A personal Japanese tutor with Anki, local Claude Code and JSON memory.
CLI-first MVP for an LLM navigation agent.
A simulated terminal, not a live agent or a preview of quitereason.ai.
Machine learning & deep learning.
From datasets and training dynamics to evaluation. Explore the Transformer, then follow a single neuron through forward pass, backpropagation and a parameter update.
Model adaptation: LoRA trains low-rank adapters; QLoRA combines a quantized base model with trainable LoRA adapters. Full fine-tuning updates the base weights. Dataset quality and evaluation matter in every approach.
Interactive illustration · not live training
Simplified Transformer, bottom to top: embeddings, multi-head self-attention, feed-forward network and output projection. Residual connections, normalization and repeated blocks are omitted. Directional signals illustrate training; no actual LLM training or inference runs here.
Predict the next token: forward pass → softmax cross-entropy → backward gradients → optimizer update.
Prefer an existing base, or continued pretraining for a domain. Scratch pretraining needs substantial justified data and compute.
Plan batches, accumulation and learning-rate schedules; track validation loss and save recoverable checkpoints.
Evaluation is continuous. Alignment is optional. Deployment feeds the next iteration.
Notes on MCP server design, reliable agents and model behavior. More to share soon.
Spot, perpetuals & prediction markets.
Low-latency architecture: sequenced ring buffers, in-memory order books, risk checks and matching engines. Explore how the market lifecycle changes without conflating matching with settlement.
Illustrative architecture — not a live exchange or Edel system diagram.
Sequenced input passes pre-match checks into an in-memory order book and matching engine, then emits execution and settlement events. Journal and replay branch from sequenced input; durability policy depends on the design.
Reference: LMAX Disruptor patterns ↗
Smart contracts, protocols & integrations.
Illustrative contract / ledger / event flow.
Contract execution connects to ledger state; emitted events feed application integrations. A conceptual view, not a claim of universal chain architecture.
At Summon: 100+ smart contracts deployed across EVM, Solana and Move, with cross-VM rewards built from one source of truth.
G7DAO/summon-api-suite ↗Non-transferable tokens for governance roles, contributor reputation, and game-state ownership. Defines a minimal standard with revocation, expiration, and on-chain attestations.
→ Read on ethereum-magicians.orgAuthored with Cody Littley and Alejandro Ranchal-Pedrosa. BUDs are an incremental authenticated data structure that scales membership proofs with per-block write volume rather than total state size — no global state trie on the critical path.
→ Read on ethresear.chComing soon
A future exploration: bringing personal agents into the physical world with LeRobot.
Concept illustration — no hardware prototype claimed.
// 07 · Games
Built games at Game7.
Coming Soon!
New worlds are loading.
Backend services, frontend applications and the infrastructure connecting them: databases, protocols, queues, cloud and AI providers.
Built games at Game7.
Open source contributions & hackathons. Public projects and merged pull requests, read from GitHub.
Latest merged pull requests across external open-source repos — rebuilt weekly.
Rust implementation of protobuf with editions support, JSON serialization, and zero-copy views
Cross AI CLI template to use for ANY development with specific skills and pipelines.
🤗 LeRobot: Making AI for Robotics more accessible with end-to-end learning
Transports, Middleware, and Networks for the Alloy project
Agent skill: stack overlapping skills — run the best candidates in parallel, score outputs, merge the result
I build across AI infrastructure, backend systems and Web3. More than 13 years of production engineering, working remotely from EMEA.
The connecting thread: determinism, latency and replay — the details that make a useful idea work in production.
