118 karma · joined May 8, 2025
Application Architect with over 10 years of experience leading the development of large-scale enterprise ecosystems. Specialist in the integration of Generative Artificial Intelligence (LLMs, RAG) and Cloud-Native architectures (OpenShift, Kubernetes). Certified Odoo ERP expert with a proven track record in the digital transformation of government and retail sectors. Focused on designing scalable solutions that optimize critical business processes.
Architecture and AI GenAI (watsonx.ai), RAG, LangChain, LLMs, Vector Databases,Agents LLM Backend & Core Java (Expert), Python (Expert), Node.js, PHP. Frontend & Mobile Angular, React, Android (Native). ERP & Business Odoo Certified (Functional & Technical), CRM, Contabilidad, eCommerce. DevOps & Cloud Docker, Kubernetes, OpenShift, CI/CD, AWS/IBM Cloud.
Programming and robotics enthusiast
Github https://github.com/davidmonterocrespo24
Medium https://medium.com/@davidmonterocrespo24
I’m working on a side project https://velxio.dev
Instead of high-level emulation, the decompiled C source is compiled directly to Xtensa machine code for the ESP32-S3. We replaced the PlayStation SDK/GPU/GTE with custom software routines, running on a DIY handheld built around a Seeed XIAO ESP32-S3 Sense.
To speed up development without constantly flashing hardware, we run the compiled firmware image directly inside a browser-based hardware simulator CLI (velxio-cli) to verify serial outputs and execution in seconds.
Having 100% matching C code makes native ports to embedded platforms completely feasible without relying on heavy JIT/interpreter glue.
The game code compiles directly to Xtensa machine code, with the PlayStation SDK, GPU, GTE, CD drive, scheduler, and controller replaced by software layers.
It runs on a Seeed Studio XIAO ESP32-S3 Sense with 8 MB PSRAM and microSD.
The post goes into the porting process, memory layout, scheduler, software rendering, debugging, and the bugs that took the most time to find
For months I've been trying to emulate the iPod Classic 6G I have in my drawer, without success. After a long time, I finally managed to get it running, although it's still somewhat rough. In QEMU, I managed to boot the original, unmodified Apple retailOS. Navigation, the disc, the screen, and the music library worked for a while, but the music player appeared partially rendered and without sound.
One thing we already know is that the firmware is from Apple and is encrypted, so there's no legal download. I have an iPod Classic 6G from 2008. With the help of Claude Code, I was able to create code to extract it using emCORE. I extracted the contents of the flash memory and decrypted the image myself.
and the boot ROM cannot be read from the running device. I confirmed the latter the hard way, and along the way, my iPod restarted several times (pressing the center and menu buttons for 5 seconds to restart it).
There is no datasheet for the SoC. Each model register that allows unlocking some part of the guest is derived from a write operation performed by the guest itself,
Code and notes, including the ~16 hypotheses I had to retract along the way
What I found interesting is that the ADC reading follows the filtered signal, so you can actually observe the analog behavior from the firmware side.
Feels like this could be really useful for teaching, especially to show how digital signals turn into analog in real circuits without needing a full lab setup
A free, open-source emulator for 19 embedded boards: Arduino, ESP32, Raspberry Pi, RISC-V , running real compiled code in your browser.
The best part: it's fully local.
No cloud dependency. No student accounts. No data leaving your network. Self-hostable with a single Docker container.
Universities and bootcamps can deploy it on their own servers and give every student access to a complete embedded development environment, for free.
I've been working on this for over a year, and just shipped v2.0 with ESP32 emulation (via QEMU), a custom RISC-V core, and Raspberry Pi 3 support that runs real Python
I recently spent some time going through MiniMind, and it’s a remarkably clean resource for understanding the modern LLM stack under the hood. It’s a minimal, end-to-end implementation of a ~25M parameter GPT-style model in pure PyTorch, designed to be trained from scratch on a single GPU
Instead of heavy abstractions, it uses straightforward PyTorch while still implementing modern architectural choices like RMSNorm, SwiGLU, RoPE, and even MoE variants. What makes it valuable is that it doesn't stop at the forward pass; the repo covers the entire training lifecycle. You can trace the data flow from tokenizer training and pretraining, right through to Supervised Fine-Tuning (SFT), LoRA, preference optimization (DPO/PPO), and distillation
It’s small enough to actually read the source code end-to-end, but realistic enough to serve as a baseline for architectural experiments rather than just a toy example.
Curious if anyone here has used this (or similar minimal codebases) to test custom architecture modifications or train highly specialized small-scale models
I'm currently testing the pipeline locally on a PC with an RTX 4060, and it's a great fit for this kind of hardware
The reason I’m specifically exploring JavaScript is because of a project I’m working on (velxio.dev), where everything runs in the browser, so having a JS/WASM-based approach would make integration much simpler
Right now I’m experimenting with a QEMU based setup and exposing it through WebSockets, but the performance isn’t great and the emulation tends to be unstable (I’ve been hitting crashes under certain workloads).
That’s partly why I’m looking into alternative approaches , either pushing more into the browser, or finding a more robust backend model.
Appreciate the suggestion!