I design and implement systems at the boundary between classical software engineering and applied AI — specifically:

• Autonomous agent loops that plan, execute, and self-correct over multi-step coding tasks
• Retrieval pipelines that ingest messy real-world documents and serve precise, ranked context to LLMs
• CLI/TUI tooling that runs locally, requires no cloud dependency, and treats safety as a design constraint

My engineering approach is shaped by one conviction: language models are only as useful as the infrastructure surrounding them. I focus on that infrastructure.

A terminal-native agent that operates on a full repository — not just individual files. Operon builds a persistent, hash-gated symbol graph across the entire codebase, then uses it to execute multi-file refactors, generate structured documentation, and answer questions about execution flow.

Key engineering decisions:

• Deterministic-first REVIEWER: verifies changes by comparing disk hash to diff memory snapshot before any LLM call — eliminates hallucinated confirmation
• CRUD fast-path: structured operations (import insertion, variable renaming, comment placement) handled via tokenize and ast without LLM — removes the most common failure mode of small local models
• 5-tier surgical diff engine: SEARCH/REPLACE patching with cascading fallbacks from exact string match through fuzzy multi-line tolerance
• Mandatory approval gate: no filesystem write occurs without explicit human confirmation; timeout auto-rejects to prevent hang
• 9-provider LLM router with hot-reload config — model switching takes effect on the next call, no restart

→ github.com/qasimio/Operon

A retrieval system designed for the documents that most RAG demos ignore: scanned PDFs, PPTX speaker notes, multi-sheet Excel files. The engineering focus is retrieval correctness under real document conditions.

Key engineering decisions:

• Custom OCR pipeline (Tesseract + Poppler) handles flattened text and image-only pages
• Hybrid search: dense vector retrieval re-ranked by a Cross-Encoder, yielding ~40% precision improvement over cosine similarity alone
• Session-isolated storage handler resolves Windows file-locking failures (WinError 32) in persistent vector stores
• Context injection optimized to reduce token usage by ~60% without precision loss

→ github.com/qasimio/MQNotebook · Live Demo

A search engine for Computer Science literature built without Lucene or ElasticSearch. The goal was to understand retrieval at the data-structure level before building RAG systems on top of it.

Key engineering decisions:

• Positional inverted index with O(1) keyword lookup via custom hashing
• Offline Indexer pre-processes corpora at build time, pushing query latency to sub-millisecond
• O(L) Trie for prefix-completion; Levenshtein distance for fuzzy matching
• TF-IDF scoring with behavioral re-ranking

The classical IR knowledge from DevShelf directly informs the hybrid search design in MQNotebook.

→ github.com/qasimio/DevShelf

A published CLI tool for organizing large, messy file collections by extension — with background watch mode, undo, deduplication, and custom category configuration. Designed for data pipeline preparation and keeping directories clean without any manual effort.

pip install foldr                          # one install, everything included

foldr ~/Downloads --preview               # see exactly what will happen
foldr ~/Downloads                         # organize (preview → confirm → move)
foldr ~/Downloads --recursive --depth 2   # include subdirectories
foldr watch ~/Downloads                   # organize now + keep watching forever
foldr undo                                # restore the last operation

Key engineering decisions:

• Background daemon architecture: foldr watch spawns a detached OS-native subprocess; the calling terminal returns immediately while the daemon runs indefinitely, using inotify (Linux), kqueue (macOS), or ReadDirectoryChangesW (Windows) — 0% CPU when idle
• Initial scan + continuous watch: daemon organizes all existing files on start, then reacts to every subsequent file creation, modification, or drag-in — no stale state, no one-time-only moves
• JSON undo system: every organize operation writes an immutable history entry; foldr undo reverses any past operation independently without requiring sequential rollback
• Zero external dependency for core output: ANSI rendering via ctypes (Win10+) with colorama fallback — no rich, no pyfiglet, no curses
• Conflict-safe moves: resolves filename collisions by appending _(1), _(2) — never overwrites
• Dry-run architecture: previews the complete I/O plan before any file is touched; directories are never modified

foldr ~/Downloads --dedup keep-newest     # remove duplicate files (irreversible — always preview first)
foldr history                             # browse past operations
foldr config --edit                       # open category config in editor

→ github.com/qasimio/foldr · pypi.org/project/foldr

• Extending Operon's symbol graph to JS/TS via Babel AST integration
• LSP server mode for editor integration
• Structured output evaluation framework for RAG retrieval quality