home whoami [ work ] blog
llm-layer-distillation

llm-layer-distillation

Hotswapping a Qwen 2.5 attention layer with a Hyena convolution.

PythonPyTorchTransformersResearch
Layer-Probing

Layer-Probing

Probing information flow, entropy, and layer-wise representations inside large language models.

PythonPyTorchHuggingFaceResearch
Agentic Researcher

Agentic Researcher

An agentic research tool that orchestrates multi-step workflows—searching, filtering, and scraping websites—to provide fact-checked answers.

PythonLangChainAgentsRAG
LinkedIn Lead Scraper

LinkedIn Lead Scraper

CLI pipeline that scrapes LinkedIn company pages, ranks outreach targets by title, scores profiles with a local LLM, and exports enriched leads to Google Sheets.

PythonCLILLMAutomation
RAG Voice Assistant

RAG Voice Assistant

A personal voice assistant application for experimenting with state-of-the-art transcription, response generation, and text-to-speech models.

PythonFastAPIRAGTTS
ReadWatch-Bot

ReadWatch-Bot

Personal media tracking PWA built with Supabase, TMDB & Jikan APIs which uses Discord bot to add/remove entries via chat commands.

JavaScriptDiscord.jsSupabaseAPIs
advent-of-code

advent-of-code

My solutions for Advent of code 2025 puzzles

Python
CryptVault

CryptVault

A desktop app for managing encrypted files with a clean vault interface, strong AES‑256 encryption, and optional VPS sync for off‑site, zero‑knowledge backups.

PythonDesktopCustomTkinterCryptographySQLCipherSecurity

browser-use

merged

A critical asyncio deadlock and Gemini integration bug were causing CI/CD crashes and API failures across the codebase. Implemented stop() logic to unblock event loops and refactored Pydantic models for strict type compliance — eliminating pipeline failures and restoring API functionality for the entire userbase. (Issues #3806, #3804)

PythonAsyncIOPydanticBug Fix

torrra

merged

A runtime failure in Python 3.10 completely blocked execution for legacy users, making the CLI unusable. Engineered a tomllib fallback strategy and used PEP 621 to conditionally inject dependencies — restoring full cross-version compatibility without breaking newer environments. (Issue #226)

PythonCompatibilityBug Fix

crawl4ai

merged

A CPU-bound chardet.detect call was blocking the main asyncio event loop causing 1+ second lags, and a tuple immutability violation in FilterChain.add_filter was triggering AttributeError crashes. Offloaded detection to asyncio.to_thread and rewrote add_filter to construct a new tuple instead of mutating — eliminating the event loop bottleneck and fixing the crash while preserving the original design intent. (Issue #1751, #1753)

PythonAsyncIOBug FixPerformance