Efrain Garay

Solutions architect and computer engineer. For nearly two decades I have been designing, building and integrating software, almost always from the same starting point: the architecture, the blueprint before the code. I draw how a system is going to work, I stay with the teams that build it, and afterwards I answer for why it was done that way. I work from Melipilla, Chile.

The last few years have been in the financial sector. I moved a banking core and its digital lending platform to the cloud: deciding what the system would look like at the end, how each piece connected to what already existed, in what order everything moved, and defending those decisions when the time came to explain them. I have also worked closely on the rollout of Chile’s Open Finance System under Fintech Law 21.521 and the CMF’s NCG 514, taking part in the regulatory sandbox: demos of the four roles the regulation defines, on concrete standards, OpenAPI 3.1, FAPI 2.0 security, OAuth 2.0 with OpenID Connect, mTLS and consent management.

Before that I was Tech Architect Leader (2021 to 2024): cloud solutions deployed through GitLab, and templates in NestJS, Spring Boot, Dataflow and Python so that no new team had to start from zero. The road before that was varied on purpose: end-to-end IoT with Arduino, AWS IoT Core, Kubernetes and a mobile app; a car-sales PWA with scraping queues; a transactional crowdfunding platform with its payment providers; and, since 2008, web development for retail and real-estate brands. The formal grounding is a degree in Computer Engineering (AIEP), with specializations in DDD, hexagonal architecture, CQRS and event-driven design.

The rest I taught myself, breaking things until I understood them. My method fits in three words: test, measure, publish. The blog works the same way: it does not republish news, it tests it on my own hardware and publishes the numbers with the failures included. Agatha came out of that stubborn curiosity: storing files inside YouTube videos and getting them back bit for bit. It ended up with a paper of its own and a public demo. Even the name has a story: it is after a dachshund whose 1-bit photo survived the compression where random noise did not.

All of it runs on a swarm of machines I own: a local GPU that generates image, voice and video, workers for audio and upscaling, a Mac that assembles, and a VPS that serves production. That is where PhantomRelay, the Agatha demo and this very site live: projects, writing and the design system I draw everything else with.

Since 2024 I have not stopped building with generative models, always as engineering of my own. It started with RAG systems over FastAPI and local LLMs. It continued with ComfyUI on my own GPU, a 16 GB RTX 4070 Ti: FLUX for images and Wan 2.2 14B quantized to GGUF for video, with RIFE interpolating frames and Real-ESRGAN upscaling to HD.

Then came the LoRAs I trained myself, not downloaded: a character LoRA from a 32-image dataset I generated and curated, trained with ai-toolkit on a rented 3090, plus face and style combinations with per-adapter weights.

With those pieces I put together end-to-end video pipelines: a script written in beats, voice cloned with Qwen3-TTS, sound effects generated with Stable Audio Open, an audio master audited to -14 LUFS, and editing synchronized to the voice. All of it spread across my swarm: one GPU generates, another interpolates and upscales, the Mac assembles and publishes.

Today that line culminates in the studio: a 13-node agentic graph in LangGraph that researches with real data, writes, directs and checks every shot with a multimodal critic, with quality gates (the transcript is compared against the script, the audio is audited by metrics) and a retry budget. It is resumable from Postgres checkpoints and operated from a web panel of my own built with FastAPI and Astro.