Final Devlog 🤖

Hey everyone!

This is my final devlog for this project.

I’ve been working on JARVIS every day, and it’s pretty much been all I’ve been doing. It’s been a challenging but rewarding project, and I’ve learned a ton along the way. Here’s a quick summary of some of the biggest technical challenges I ran into.

Dual Database Architecture

One of the hardest parts was managing two databases at the same time. Every message gets saved twice: SQLite stores the raw chat history, while ChromaDB stores vector embeddings for semantic memory. Every write has to handle the possibility of ChromaDB being unavailable, and every read has to combine recent chat history with similarity based memory retrieval. If either system breaks, JARVIS still works, but it loses its long term memory.

Replacing PyTorch with ONNX

PyTorch turned out to be a huge problem for deployment. The sentence-transformers package pulls in roughly 1.5GB of CUDA libraries, while most free hosting platforms only provide 512MB of RAM. I switched to ChromaDB’s ONNX embedding runtime instead, which reduced the model to around 30MB. Getting ChromaDB to consistently use ONNX instead of falling back to PyTorch took a lot of configuration, and if the embedding function changes between deployments, the entire vector database has to be rebuilt.

Authentication

Even though JARVIS is a personal assistant, I still wanted authentication to be done properly. User accounts use bcrypt password hashing, username validation, password validation, duplicate username checks, and identical login error messages to prevent username enumeration. It’s definitely overkill for a project like this, but it was a great learning experience.

Overall, this project taught me a lot about AI systems, deployment, databases, authentication, and building software that actually has to work reliably. I’m really happy with how JARVIS turned out, and I’m excited to keep improving it even after this project is over.

The boring stuff is all done now.