Hi, I'm Adeniji, here's what I'm building

Knight Eterniti: OCR + RAG on AMD ROCm and Xvimo

I am building an OCR + RAG pipeline on AMD GPUs by running OCR (Tesseract or PyTorch-based models like TrOCR/Donut) to extract and chunk text, encoding chunks with a ROCm-compatible PyTorch embedding model, storing embeddings in a vector store (FAISS on CPU for small scale or Milvus/Weaviate for production), retrieving top-K chunks for a query, and feeding those into a generator LLM running on ROCm.
In addition to the above, I built xvimo to reduce money-loss from scams in regions where WhatsApp is the primary channel for payments and checks. xvimo runs as a WhatsApp-accessible agent that people can forward suspicious messages to; it performs rapid checks (link analysis, scam-pattern detection, payment-request heuristics) and returns an actionable verdict and short, safe advice (e.g., “Do not send money — likely fraud; verify with bank and sender via a different channel”).
What made it “impossible” to some was designing an agent that:

1. works within WhatsApp’s constraints (low bandwidth, short text flows, no heavy client installs),
2. prioritizes privacy (no raw message storage; ephemeral checks), and
3. produces usable, low-friction outcomes for non-technical users. We’ve validated the flow with pilot users and simulated datasets showing meaningful precision on common scams while keeping false positives low enough that people follow recommendations.

Link to check it out: xvimo.onrender.com

It's nice joining everyone.