I specialize in building performance-first developer tooling, local-first machine learning pipelines, and hybrid AI systems optimized for low-latency execution, hardware efficiency, and absolute data privacy. My goal is to shift complex computational workloads back to client hardware through optimized architectures, lightweight compilers, and custom client runtimes.

• 🤖 Agentic AI & Coding Assistants: Local-first CLI agents with zero-trust execution sandboxing, AST-memoized 60 FPS terminal streaming (React Ink), Model Context Protocol (MCP v1.29 SDK), and multi-provider offline inference (Ollama, LM Studio).
• 🧠 Deep Learning & Transformers: Autoregressive decoder-only Transformer design (causal self-attention, rotary positional embeddings (RoPE), KV-caching, FlashAttention) trained from scratch.
• 🔍 Retrieval-Augmented Generation (RAG): Intent-driven multi-agent workflows (LangGraph), high-performance vector databases (Qdrant, FAISS), and Reciprocal Rank Fusion (RRF) for hybrid retrieval.
• 🛠️ Systems & Developer Tooling: Abstract Syntax Tree (AST) parsing with WebAssembly Tree-sitter, VS Code extension development, and custom Electron/Code-OSS IDE compilation.
• 💻 Desktop & Native Wrappers: Lightweight native GUI applications using PyWebView, FastAPI, Monaco Editor integration, and sub-process compiler integration.

Enterprise-grade, local-first agentic CLI coding assistant — uniting private offline LLMs and 11 multi-cloud AI providers with zero-leak security and 10 autonomous workspace tools.

• Autonomous Agent Loop & Universal Parser: Intercepts and executes function calls across all formats (Groq/Llama 3 inline, Anthropic XML, OpenAI/Ollama XML, and Markdown JSON) with adaptive token budgeting and plan-and-apply dual execution modes.
• Enterprise Zero-Leak Security Vault: Enforces machine-bound AES-256-GCM credential encryption (0600 permissions), OS Keychain sync, strict execFileDirect command isolation, realpath symlink containment, and IPv4-mapped IPv6 SSRF defense.
• 60 FPS Reactive Terminal TUI: AST-memoized streaming at 60 FPS eliminating terminal redraw jitter, non-blocking cursor movement, multi-phase Tab/Shift+Tab autocomplete, and multi-line clipboard pasting.
• 11 Inference Engines & MCP Ecosystem: Native offline local runtimes (Ollama, LM Studio) + 9 frontier cloud providers (NVIDIA NIM, Anthropic, Groq, DeepSeek, OpenAI), Model Context Protocol (MCP v1.29 SDK) server tools, and persistent project memory (.agentmemory).
• Stack: TypeScript 5 • Node.js 20+ • React (Ink 7) • NVIDIA NIM • Ollama • LM Studio • Anthropic • Groq • MCP SDK • AES-256-GCM

Multilingual GPT-style autocomplete model trained from scratch supporting Python, C, and Java.

• Architecture & Design: Integrated Rotary Positional Embeddings (RoPE), GeGLU gating, and soft logit capping into a custom causal decoder framework.
• Optimization: Enabled FlashAttention (SDPA) and dynamic KV-caching to accelerate local inference speeds on consumer GPUs.
• Stack: PyTorch • Transformers • FlashAttention • BPE Tokenizer • DDP

Fine-tuned academic study tutor built on top of Qwen-2.5-7B-Instruct, optimized for Colab training & local GGUF running.

• Dataset Pipeline: Engineered a custom 5-layer dataset pipeline spanning ~480K instruction-tuning samples and 15K preference pairs.
• Training & Quantization: Configured a multi-phase pipeline executing SFT and DPO via Unsloth, quantized to GGUF format for low-latency running locally with Ollama/llama.cpp.
• Stack: PyTorch • Qwen-2.5 • Unsloth • DPO / SFT • HuggingFace • GGUF • Ollama

Privacy-first local search engine transforming directories and Obsidian vaults into context-aware systems.

• Ingestion & Sync: Engineered a debounced filesystem watcher parsing markdown links and headers in real-time.
• Retrieval Engine: Fused vector search scores with SQLite structural centrality metrics using Reciprocal Rank Fusion (RRF) for highly contextual answers.
• Agent Workflows: Designed self-correcting agent state graphs verifying prompt retrieval context to prevent LLM hallucinations.
• Stack: Python • LangGraph • Qdrant • Ollama • SQLite

A custom Windows distribution of VS Code (Code - OSS) with a built-in AI autocomplete companion.

• Built-in Extension: Developed a custom completion provider utilizing Fill-in-the-Middle (FIM) prompt formatting.
• Latency & Caching: Optimized client-side response paths through request debouncing and predictive local caching.
• Stack: TypeScript • Electron • Node.js • Ollama

A native desktop IDE for Ubuntu executing client-side compilation and sub-process GCC checks.

• Features: Developed client-side compiling, sub-process GCC diagnostics, and real-time gutter diagnostics.
• Visuals: Features interactive AST analysis and local visual memory allocation charts.
• Stack: Python • FastAPI • Monaco Editor • PyWebView

Real-time code visualizer rendering DSA recursion trees and UML diagrams as you type.

• AST Parsing: Developed an offline-first parsing engine running WebAssembly-compiled Tree-sitter AST queries in-browser.
• Visualization: Transformed code structures dynamically into interactive Mermaid.js diagrams.
• Stack: TypeScript • WebAssembly • Tree-sitter • Mermaid.js • VS Code Extension API

Local-first knowledge base assistant ingesting documents into a vector database with a central command UI.

• Ingestion: Ingests .pdf, .txt, and .md files dynamically using local embedding models into a FAISS vector store.
• Interface: Features a futuristic browser console with MathJax LaTeX rendering, marked-down contexts, and a live search telemetry stream.
• Stack: Python • FastAPI • FAISS • Gemini API • Tailwind CSS

Explore 23+ projects, research tools, and simulators on my Live Portfolio.