Models hallucinate. Systems shouldn't. I don't build model wrappers. I build the retrieval pipelines, autonomous loops, and architectural constraints that force LLMs to be reliable, safe, and grounded in reality.

Engram — Next.js FastAPI pgvector RAG A local-first knowledge OS. Enforces strict, evidence-bound retrieval so the LLM returns "not found" instead of hallucinating. Workspace-isolated, async-ingested, provider-agnostic.

Operon — Python AST Textual ReAct A terminal-native coding agent that reasons across full repositories. Verifies filesystem state deterministically before trusting LLM output. Mandates human approval before any write.

FOLDR — Python PyPI Daemon Cross-platform file automation CLI. Zero-CPU background daemon with git-style filesystem undo. 7,000+ downloads.

MQNotebook — LlamaIndex Tesseract Streamlit Enterprise RAG for messy reality. Custom OCR pipeline for scanned PDFs and multi-sheet XLSX. Cross-encoder reranking yields a 40% precision lift over baseline.

DevShelf — Java Information Retrieval A search engine built from first principles. No Lucene. No Elasticsearch. O(1) positional inverted index with sub-millisecond query latency.

• FlyRank AI — Backend & AI Engineering Intern. Building application layers around model-driven workflows.
• Arch Technologies — ML Engineering Intern. BERT fine-tuning and containerized local inference deployment.
• Sukkur IBA University — CS Community Lead. Managing a network of 300+ students; reviewed 50+ technical projects for architectural integrity.
• ConcertHub — Collaborative Next.js deployment. Demonstrating architecture and integration outside solo environments.