Rashaad N Mohammed, AI/ML Engineer RRAASSHHAAAADD

A local tool for actually looking at the vectors inside a Chroma database: PCA-reduced and rendered as an interactive 2D, 3D, or unit-sphere view. Companion to my write-up on embeddings.

• Python
• FastAPI
• ChromaDB
• Plotly.js

View, download and manage files across multiple linked Google accounts from one page, with session-scoped caching and read-only Drive OAuth. Shipped alongside: a public ad-booking form, an admin approval panel and live ad-slot rendering on Prisma/PostgreSQL, with encrypted tokens for every account.

• TypeScript
• Express
• PostgreSQL
• Prisma
• Google OAuth

A privacy-first assistant on local Mistral-7B that scans Gmail, filters automated senders and drafts replies to high-priority mail, with nothing leaving the machine. Runs on an i5 with no GPU.

• Python
• MCP
• Mistral-7B
• ChromaDB

Cleaned a resume dataset into role, experience and keyword targets, then fine-tuned Qwen2.5-1.5B-Instruct with LoRA to pull structured tech-stack profiles from raw resume text. Training on Modal through a repeatable dataset-to-checkpoint pipeline.

• Python
• PEFT / TRL
• Qwen2.5
• LoRA
• Modal

A fully offline voice assistant on a Raspberry Pi with sub-second latency: speech-to-text, the LLM and text-to-speech all on-device, RAG memory for user-taught facts, 4-bit quantized to run under 2GB of RAM.

• Python
• llama.cpp
• RAG
• Raspberry Pi

A full-stack lead-research product: a FastAPI layer, LangChain agents for company research and contact discovery, and a Pandas ingestion layer for messy client inputs. Surfaced 3 to 5 verified contacts per company on average.

• Python
• FastAPI
• LangChain
• LLM API
• Pandas

A Deep Q-Network never given a rulebook: +10 for eating, −10 for dying, across 85,000 games. Includes the checkpoint bug that silently rewound training. Full story in the write-up.

• Python
• PyTorch
• Reinforcement Learning
• DQN