Software Engineer

Alexis Alulema

Software Engineer · Builder

Software Engineer with 20+ years turning complex problems into scalable solutions — across software, cloud, AI and hardware.

Read the blog View projects
MCP Talk to this site from your AI assistant

Try it live

Real AI projects you can spin up on demand — each one gets its own ephemeral environment.

See all projects →
Live demo

Trip Planner (Chain of Agents)

Type a trip request and watch a chain of specialised agents hand the work along in real time until they deliver a day-by-day itinerary checked against your budget. A LangGraph orchestrator owns the control flow; a local Qwen 2.5 writes only the prose, while typed decisions and plain arithmetic handle the rest. When the plan goes over budget, a conflict-resolution agent renegotiates it without a human. Local LLM (Qwen 2.5 via Ollama) and no API keys; live weather, places and exchange rates come from open data sources.

PythonFastAPILangGraph
Try the live demo →
Live demo

Agentic Racing

Six identical cars race a closed circuit in the browser. Each car is a two-tier agent: a pilot that drives every frame, and an independent LLM team boss that reads live telemetry and radios back a strategy — attack, defend, conserve — F1 team-radio style, on screen. The race never waits on the model: a fixed heuristic takes over the instant it is slow, off, or replies invalid JSON. Fully self-hosted: a local llama3.2:3b behind a proxy with its own circuit breaker, rate limit and concurrency gate.

Unity WebGLPythonFastAPI
Try the live demo →
Live demo

RAG Chatbot over Blog Posts

An interactive retrieval-augmented generation chatbot that answers questions using my blog posts as a knowledge base. Fully self-hosted: semantic vector search (pgvector) with multilingual embeddings and a local Qwen 2.5 LLM served by Ollama — no external LLM APIs.

PythonFastAPIpgvector
Try the live demo →
Series · 5

Reinforcement Learning from scratch

Five articles that teach RL using a real project — Agentic Racing — as the lab: from what an agent is, to PPO, imitation and how to combine them.

Start with part 1 →
  1. Part 1How Does an Agent Learn? Teaching a Car to Drive
  2. Part 2How Do We Improve a Policy Without Breaking It? PPO Explained from Scratch
  3. Part 3Can We Teach by Imitation? Behavioral Cloning Explained from Scratch
  4. Part 4Can We Learn to Behave Like an Expert? GAIL Explained from Scratch
  5. Part 5Can We Combine Imitation and RL? BC, GAIL, and PPO as One System

Latest posts

See all posts →
10 min read

The MCP server that waited almost two months (and shipped in one night)

This blog is now also a read-only MCP server: any AI assistant can search and read the posts and check the demos, always citing the original URL. I had the idea in September and almost every piece already existed, but the demos, especially Agentic Racing, came first. When I finally finished the spec, the server was in production that same night. This post covers why it took so long, what was actually missing, and how search turned out: recall@5 of 1.00 with the post's own words and 0.40 with paraphrases.

mcpagentsllmraginformation-retrievalpython
15 min read

Harness²: an agent that improves its own instructions (and a judge that always said yes)

I turned the core idea of Harness², from Google Research, into a live demo: improve an agent by changing its execution harness, not the model, in a Propose → Probe → Compose loop. With 1.5B and 2B models on CPU and the same model acting as judge, all 26 comparisons I recorded accepted the proposed change, 18 of them with exactly the same scores, and the judge rewarded answers with the arithmetic backwards. This post covers what I built, how I measured it, why an optimizer has no way of knowing whether it improved without an independent verifier, and under what conditions I'd use Harness² in my next projects.

agentsllmllm-evaluationprompt-engineeringollamapythonfastapi
16 min read

The Dreaming Agent: Three-Level Memory with a Small Model (and the Naive RAG That Beat It)

I turned LightMem's memory architecture into a demo: an agent that stores facts as it talks and that, when you ask it to "dream", consolidates its memory —merging, updating, promoting and forgetting— as a visible diff. Fully local, with 3B and 7B models. I measured it against a naive RAG with three runs per condition, and it lost: 79 vs 88 % with 7B, 58 vs 93 % with 3B. This post covers what I built, where it breaks, and why, with small models, the best memory turned out to be the one that rewrites the least.

agentsllmmemoryragollamapythonfastapiembeddings