CyberNovas

A tiny ~10K-parameter LLM router that learns which open-source model (deepseek-v4-pro / glm-5p2 / kimi-k2p6 via Fireworks) should answer each question and in what role, trained by evolution (sep-CMA-ES).

GPT-2 train from scratch in pure NumPy, verified against PyTorch. I have written a blog explaining all the math. Check out the website.https://x.com/HarshalsinghCN/status/2055154124535287829?s=20

Preference-supervised naturalness scorer for modern neural TTS . best way to measure naturalness

Clearing the nanoGPT speedrun's 3.28 val-loss target on one H200 by stacking the Aurora optimizer and Token Superposition Training (TST) on an untouched Transformer.

No description.

SOTA Hinglish (89M) code-switching TTS, fully in the browser on WebGPU

change handwritten pdf into typed beautiful pdf

No description.

Does HRM-style hierarchical recurrence beat a matched-parameter transformer on audio tokens? A head-to-head experiment.

Self-improving LinkedIn content engine — multi-agent simulation, EMA weight calibration, 10-persona debate threads, cron automation. Built for Indian SMBs. visit humanless.site

linkedin matchmaking dating using llms

An Agent Skill (Claude Code + Codex) that turns your X/Twitter bookmarks into a local, searchable, plain-markdown knowledge base with a command-palette UI. No login, no API keys, all on disk.

MV3 Chrome extension that captures your Twitter attention log — tweets, sessions, dwell time — into a local SQLite DB with daily markdown exports.

mini alphafold from scratch

Rag implementation from scratch without any framework like langchain or llamaindex

chatbot for tenants and landlords

Hinglish TTS: uses 100% synthetic, two-gate-verified corpus, with a convention-robust bin-aware filter and paired objective eval. Model on HF: harrrshall/xtts-v2-hinglish-synthetic

https://harrrshall.github.io/.

Open-ended confound-controlled research: text-derived prosody as a complementary signal for writing quality (slop-free, diversity-preserving)

AI & semiconductor hardware research pipeline 2025-2026

A multi-agent research skill for Claude Code & OpenAI Codex that hunts genuinely non-consensus findings in any topic domain, with independent market-pricing verification.

React PCB Viewer for Circuit JSON

Validity and Design Rule checks for Circuit JSON

Brand counterfeit detection — Brand Discovery Agent + classifier with perceptual image hashing.

chat with website

This code generates a podcast from a script in a TXT file. It uses a lightweight model (Kokoro-82M) that can run on any laptop.

thinknote

No description.

No description.

collection of scripts to automate random things

Heyy It's Harshal

No description.

Open-Podcast is a Python project that leverages the Kokoro library to generate audio podcasts from a script. It provides a flexible framework to define different speakers, their voices, and generate natural-sounding audio segments with appropriate pauses.

A simple way to make attendace just by uploading pictures of classroom, require no signup,no setup nothing

No description.

making personal computer more "personal"

revolutionizing college and study exams

turn podcasts into engaging blogs

arjun , ask to the shri krishna

count calories using camera

remove physcial digital barrier

turn your research papers into working code

Implementation of full fledeged tranformer(ie. attention is all you need paper) in c, for learning purpose inspired by @andrew_karpathy

simple rag chat with youtube video

this is fun projectr in javascript that enable phone to talk to the pc, and execute some defined command

For the sake of learning, I'm implementing LLAMA from the ground up in my own way

speaker what you are doing using gpt4-vision

with the help of gpt-vision it will be label all the image and sort then them using my python script

Here is my implementation of Andrew Karpathy's neural network series in C to test my understanding.

Neural Networks from Scratch in various programming languages