Hello.
I build things
that I'm passionate about.

Trained a 146.8M-parameter decoder-only GQA transformer for 30,000 steps on a blended corpus of FineWeb, OpenWebMath, and Gutenberg/BookCorpus data using bfloat16 AMP and gradient checkpointing. Built the end-to-end training pipeline for dataset ingestion, token packing, distributed training, checkpointing, prompt-based evaluation, and Vertex AI deployment on 2x NVIDIA L4 GPUs. Processed 27.3B packed tokens and trained on an estimated 4.42B tokens, reaching a best validation loss of 3.2709

Supports local llama.cpp (llama-server) GGUF model workflows with model selection/loading, plus custom endpoint chat with real-time token streaming. Supports resilient runtime fallback between pipeline modes, knowledge ingestion (chunking, embeddings, semantic search), and feedback history export in OpenAI-style JSON/JSONL format.

Fast CLI tool to summarize, compare, and visualize the graph structure of Hugging Face open-weight LLMs, streamlining model evaluation. Core engine built in Rust for high-performance comparisons, summarization, and graph generation. Python bindings handle deep inspections like downloading models and local analysis, alongside an interactive TUI.

This is where I started trainig models from scratch. I did pre-training and supervised fine-tuning (SFT) on three model sizes, Basic (7.29M), Mid-Level (12.39M) and Quality (40.73M) params. Best val loss hit 1.86 after fixing weight init, packed tokenization and cosine lr decay.

A model about myself. I fine-tuned SmolLM with LoRA and QLoRA on a 2K sample dataset that I curated and reviewd manually. Covers my bio, career, skills, and some refusal + safety categories.
see your bottom right corner please →

High-performance offline voice-to-text desktop tool built with Tauri v2, Rust, and TypeScript. Robust backend audio pipeline in Rust using cpal for background microphone capture, adaptive silence detection, and whisper.cpp for fast, local transcription. Seamless hands-free text auto-pasting across macOS, Windows, and Linux via OS-specific global hotkeys and clipboard APIs.

Fast, cross-platform CLI tool to securely expose local server ports to the internet. Uses Go Routines for high-concurrency, low-latency persistent WebSocket connections. A simple, self-hosted alternative to ngrok for local-to-public tunneling without complex firewall or NAT configurations.

API Gateway for microservices architecture using Spring Cloud Gateway, integrated with Netflix Eureka for service discovery. Configured advanced resilience patterns like automatic retries, circuit breaker (Resilience4j), and fallback routes. Centralized routing, load balancing, and Redis-based rate limiting to protect backend services from abuse or overload.

All-in-one AI platform designed to empower businesses with content creation, SEO optimization, and marketing automation capabilities. Leverages advanced AI to generate high-quality content and automate marketing workflows.

AI-powered talent platform designed to connect employers with skilled professionals. Offers advanced matching algorithms to streamline hiring and integrates robust collaboration tools for seamless project management.

Microservices-based web application to reduce food waste by connecting donors and recipients. Built with Express.js for user, post, like, comment, and claim services and FastAPI for real-time notification services.

Personal automation utility for browser and command line tasks. Multi-browser launch, command runner, and bookmark importer. Custom browser mappings for URLs and imported bookmarks from Chrome, Brave, Edge with auto-detected profile handling.

Clipboard history manager with Ctrl + Alt + Z for instant access. Automatically detects whether copied content is regular text or source code. Opens generated files directly in Visual Studio Code for seamless editing.

Reduces boilerplate code when working with Reactive Forms in Angular, handling form initialization, validation, and error management. Easily integrates to display meaningful error messages for invalid fields.