iiinova — where your digital self keeps evolving

Hemanth HM @gnumanth.bsky.social · 24/09/2026

A decade ago, I started a tiny GitHub project to list functional programming jargon. Today, it has over 18k stars, has been translated into 13+ languages, and is getting a fresh new look! Interactive graph and a detailed information sidebar. hemanth.github.io/functional-p...

000

Hemanth HM @gnumanth.bsky.social · 15/09/2026

Using OpenAI's GPT‑Live‑1 and Gemini's Live APIs I baked Phonica, an app that helps in pronouncing random words across 22 languages. P.S: It is bit of challenge today to achieve a similar state with FOSS models, will keep pawing on the side. h3manth.com/ai/phonica/ #ai #llm #language

120

Hemanth HM @gnumanth.bsky.social · 09/09/2026

Baked Statewise with OpenAI's gpt-6-astra: interactive map, facts, search & 3 quiz modes. Not a one-shot. Great at responsive UI, quiz logic, and git hygiene, but required constant follow-ups for UI quirks & demos. Capable, but a high-maintenance collaborator. h3manth.com/fun/statewise/

000

Hemanth HM @gnumanth.bsky.social · 08/09/2026

Baked browser-llm-fit: check if an AI model fits your browser hardware before downloading weights. Probes WebGPU limits, shader-f16 & RAM so tabs don't crash. Try it: h3manth.com/ai/browser-l... npm i browser-llm-fit #ai #llm #web

h3manth.com

browser-llm-fit — Check if AI models fit your browser

Auto-detect WebGPU hardware, WASM SIMD, and buffer memory limits to determine which AI models run client-side inside your browser tab.

110

Hemanth HM @gnumanth.bsky.social · 03/09/2026

Anthropic open-source blueprint: commerce agents. I added voice layer, for fun. Two agents, define once: - Shopping: search, compare, cart, policy Q&A, remember prefs. - Merchant: explain perf, edit listings, react to inventory, price/promote.

200

Hemanth HM @gnumanth.bsky.social · 01/09/2026

Every argument has an underlying geometry... I baked "logically" an interactive research workbench and reference compendium for formal logic, philosophical razors, cognitive biases, and automated dialectic argument analysis. h3manth.com/fun/logically/ #Philosophy #Logic #CriticalThinking

020

Hemanth HM @gnumanth.bsky.social · 31/08/2026

Node's REPL now has basic syntax highlighting support. v26.8.1 🤓 #node #javascirpt

000

Hemanth HM @gnumanth.bsky.social · 28/08/2026

Real speech-to-speech + accurate tool calling + low latency is a hard combo. PipeCat's phone-call LLM does it via a cascading pipeline GPT-level tool accuracy, low latency, cheap.Turns out you don't need native S2S to ship good voice agents. #voice #llm #ai #agents

210

Hemanth HM @gnumanth.bsky.social · 25/08/2026

Introducing jeopardygithub.com convert any GitHub repo into a multiplayer classic jeopardy board!

000

Hemanth HM @gnumanth.bsky.social · 19/08/2026

RAG should be zero ceremony; piragi hits V1. Zero config, built-in vector store, streaming loader, thread-safe, and 3-5x faster parallel retrieval. Playground: hemanth.github.io/piragi/ pip install piragi from piragi import Ragi text, citations = Ragi("./docs").ask("How to deploy?") #AI #ML

010

Hemanth HM @gnumanth.bsky.social · 13/08/2026

Skills over MCP. People called MCP dead when Skills shipped. Wrong question! I filed an RFC for remote, URL-based skills instead. Feb 2026: MCP Interest Group formed. Apr 14: SEP-2640 charter locked. Not merged yet, but the direction's set. Demo: skills-over-mcp.h3manth.com

110

Hemanth HM @gnumanth.bsky.social · 10/08/2026

Meta's Muse-Glimmer-30B, quantized to ~16GB, on llama.cpp, analyzing XKCD INC. comics! Baked a small agentic demo: - model picks a random xkcd comic - calls a tool to fetch it - actually looks at the image, then explains the joke. #LLM #opensource

030

Hemanth HM @gnumanth.bsky.social · 08/08/2026

piragi gets ONNX and Qdrant + TurboQuant support! Faster embedding: Ragi("./docs", config={"embedding": {"backend": "onnx"}}) Less mem for vector storage: Ragi("./docs", store=QdrantStore(quantization="turboquant")) pypi.org/project/pira... #RAG #Python #ONNX #Qdrant #TurboQuant #OpenSource

030

Hemanth HM @gnumanth.bsky.social · 07/08/2026

# Load a single plugin package directory agent = await Agent("assistant").with_plugins("./plugins/bigquery-reports") # Or load multiple plugins at once agent = await Agent("assistant").with_plugins([ "./plugins/reports-plugin", "./plugins/data-agent-kit" ])

000

Hemanth HM @gnumanth.bsky.social · 07/08/2026

agentu gets agent plugins capabilities! Google, Amazon, MSFT, Cursor, OpenAI & Vercel released Agent Plugins 1.0.0 Basically standardizing SKILLmd + mcp.json into one portable folder. Reminded /me of pre-2011 Node before package.json. hemanth.github.io/agentu-codel... #ai #llm #agents #python

210

Hemanth HM @gnumanth.bsky.social · 06/08/2026

I always wanted and was baking something like this on the side, noticed this in Meta Facebook's Muse Code CLI. STT by default in you terminal harness is a pretty solid feature! #ai #agents #cli

100

Hemanth HM @gnumanth.bsky.social · 05/08/2026

Baked agenthood: a Robinhood trading CLI powered by agentu and Robinhood MCP. pip install agenthood Indeed we have a REPL and auto token refresh. github.com/hemanth/agen... #python #agentic #ai #opensource

030

Hemanth HM @gnumanth.bsky.social · 31/07/2026

Baked Semantra: semantic search fully in-browser. No backend, no API key. Snowflake Arctic-Embed-S (30MB) + WebGPU cosine ranking, cached after first load. npm install semantra "machine learning" → 83% match, zero server infra. hemanth.github.io/semantra/ #AI #WebAI #LLM

010

Hemanth HM @gnumanth.bsky.social · 29/07/2026

MCP 2026-07-28 is out and it's a big shift. The protocol is now stateless! Updated mcphost.link to support it, while staying backward compatible with older servers. #MCP #OpenSource #WebStandards

010

Hemanth HM @gnumanth.bsky.social · 28/07/2026

we can also enable speculative decoding

000

Hemanth HM @gnumanth.bsky.social · 28/07/2026

@developer.chrome.com browser gets Gemma 4 by default now, no more nano. Also, looks like we can load an external model, that could be optimized for inference, interesting...

100

Hemanth HM @gnumanth.bsky.social · 17/07/2026

You can also programmatically clone any voice, with just three lines of code: import clone from 'clone-voice'; const voice = await clone('./my-voice.wav'); const wav = await voice.speak('Hello from my cloned voice!'); npm install clone-voice -> www.npmjs.com/package/clon... #llm #voice

000

Hemanth HM @gnumanth.bsky.social · 16/07/2026

Introducing Pocket Voice: Clone any voice in your browser! Record your voice, clone it, type some text, and hear it speak. Runs entirely in your browser via ONNX and WebAssembly. No server. No API keys. Your data stays on your machine. Try it -> h3manth.com/ai/pocket-vo... !

021

Hemanth HM @gnumanth.bsky.social · 15/07/2026

Introducing Lipi, a voice first Kannada editor! Speak in Kannada and watch your words stream onto the page in real time. Type in English and it converts to Kannada. Edit, copy, save. h3manth.com/ai/lipi/ P.S: Kannada is a Dravidian language, ~55million speakers. #AI #LLM #ondevice #Kannada

h3manth.com

ಲಿಪಿ — Kannada Voice Editor

Speak Kannada, stream text, edit, save. Powered by Whisper AI.

000

Hemanth HM @gnumanth.bsky.social · 14/07/2026

Great question! We now have write-ahead checkpointing (WAL pattern) for mid-tool-call crash recovery!

000

Hemanth HM @gnumanth.bsky.social · 13/07/2026

Sanskrit fuses words together and conjugates verbs across 3×3×10 grids. Most NLP tools choke on this. Shipped two zero-dep Python packages: sanskrit-tokenizer: reverse sandhi splitting dhaturupa: verb conjugation + reverse lookup. #NLP #LLM #Tokenization #OpenSource #Python

040

Hemanth HM @gnumanth.bsky.social · 09/07/2026

agentu 2.0 is out! What changed? LanceDB with pluggable storage backends, added sandboxed tools, MCP streamable-HTTP, checkpoint/resume, typed output + OTel, and refactored the Agent god-object into mixins. pip install agentu #Python #AI #LLM #OpenSource

250

Hemanth HM @gnumanth.bsky.social · 23/06/2026

As a part of Google Antigravity's Agentic Architect Sprint I baked agy-mux: tmux for AI agents using agy CLI. Start a session → detach → resume from anywhere. Cloud VM, real PTY, WebSocket streaming. One command to deploy. Try it -> npx agy-mux #AgenticArchitect #GoogleAntigravity

010

Hemanth HM @gnumanth.bsky.social · 09/06/2026

Baked stream-slots: Out-of-order HTML streaming via Declarative Partial Updates. Define named slots, fill them in any order, and the browser renders content as it arrives. No client-side JavaScript required. npm i stream-slots 👇 #web #javascript

260

Hemanth HM @gnumanth.bsky.social · 05/06/2026

How well do you actually know the codebases you depend on? Let AI quiz you! Built a minimalist sandbox that analyzes public repositories and generates deep architectural assessments. Would you like to try it? ;)

000

Hemanth HM @gnumanth.bsky.social · 03/06/2026

form-to-chat: turn any HTML form into a step-by-step conversation. JS library, web component, Chrome extension, and an agent protocol for LLM function calling. Zero deps. npm install form-to-chat #javascript #ai #opensource

240

Hemanth HM @gnumanth.bsky.social · 29/05/2026

I baked a geo-guesser-like app and slapped the Gemini Real Time API over it! 🤓 Makes the game more fun and interesting. I have hosted it but have limited tokens on the API. Would you sign up for something like this?

000

Hemanth HM @gnumanth.bsky.social · 13/05/2026

LLMs can "learn", maybe. That doesn't mean we offload learning. I built paw-on-es6 over a decade ago a hands-on way to pick up ES6. It did its job. Now I'm putting out ESlabs. It covers every major feature from ES2015 through ES2025. eslabs.web.app #JavaScript #TC39 #OpenSource

eslabs.web.app

ESLabs - Learn Modern JavaScript by Doing

63 hands-on lessons covering let/const, async/await, optional chaining, Set methods, and more. Run real JS in your browser.

030

Hemanth HM @gnumanth.bsky.social · 21/04/2026

agentathon.dev Introducing Agentathon! Hackathon for AI Agents. Your agent enrolls, codes, and ships. Gets judged by another agent. Send your agents to the Agentathon! 👇 #AI #agents #hackathon

agentathon.dev

Agentathon: AI Agent Hackathon

A hackathon where AI agents compete autonomously. Build an agent, point it at the API, and let it enroll, pick a category, write code, and get judged.

000

Hemanth HM @gnumanth.bsky.social · 20/04/2026

Baked webml-kit. Framework-agnostic utilities for loading and running ML models in the browser via WebGPU/WASM. npm install webml-kit What's in it: → Device detection → Token streaming → GPU recovery → Cache visibility → Hub search #WebGPU #WebML #JavaScript #WebAI

010

Hemanth HM @gnumanth.bsky.social · 16/04/2026

Introducing Echo Mono: anonymous voice conversations with people who share your interests. It's peer-to-peer (WebRTC), so audio never touches a server. You get an icebreaker question to get the conversation going. Try it -> echomono.com P.S: Running on not so strong boxes, will scale if needed.

000

Hemanth HM @gnumanth.bsky.social · 17/03/2026

Baked XKCD Reverse Lookup using Google DeepMind's gemini-embedding-2-preview multimodal embedding model 🤓 github.com/hemanth/xkcd...

000

Hemanth HM @gnumanth.bsky.social · 04/03/2026

200 milliseconds. That's the exact window your brain gives someone to reply before a silence goes from normal to weird. It's not an arbitrary UX rule, it's literally just biology. I wrote down some thoughts on what that actually looks like under the hood 👇 h3manth.com/scribe/voice... #AI #Voice

h3manth.com

The 200ms window: why voice AI latency is a biology problem | Scribe by Hemanth

Human conversation runs on a 200ms biological clock. Here's what that means if you're building voice AI.

001

Hemanth HM @gnumanth.bsky.social · 03/03/2026

Baked an on-device visual narrator by running Qwen3.5-0.8B on WebGPU 🤓 It can describe, analyze, or extract text from any pasted or uploaded image, all without your data ever leaving your machine. Try it 👇 h3manth.com/ai/visual-na... #AI #WebAI #LLM

011

Hemanth HM @gnumanth.bsky.social · 22/02/2026

Taalas HC1 implements model weights directly into silicon logic, bypassing the instruction-fetch overhead of traditional GPUs. Hitting 17k tokens per second! "Mask-ROM" layers makes the chip tied to a particular model version with no flexibility for any changes after tapeout. #AI #Hardware

000

Hemanth HM @gnumanth.bsky.social · 21/02/2026

I baked Samvada a fully Indic-stack AI Counselor that uses Sarvam's models for almost realtime inference 🤓 Audio → VAD -> STT (saarika:v2.5) → Text → LLM (sarvam-m) → Text → TTS (bulbul:v2) → Audio P.S: Will try to host, but need to get the compliance right. #IndicAI #SarvamAI #VoiceAI

000

Hemanth HM @gnumanth.bsky.social · 17/02/2026

Built ucpify because every commerce backend starts the same way: products, checkout, payments, auth. Now it's one JSON file: → npx ucpify serve config.json → 13 endpoints, Stripe + PayPal, OAuth 2.0 → Node.js and Python That's it. No framework. Just a config. h3manth.com/scribe/ucpif...

h3manth.com

ucpify: JSON → Commerce Server in Seconds | Scribe by Hemanth

Generate a fully UCP-compliant commerce server from a single JSON config. Stripe, PayPal, SQLite, admin dashboard, all wired up. npm install ucpify or pip install ucpify.

000

Hemanth HM @gnumanth.bsky.social · 14/02/2026

Built webmcp-bridge to map remote MCP tools into WebMCP in the browser. Connect MCP server -> discover tools -> register with navigator.modelContext.provideContext() -> invoke from browser AI (Prompt API) surfaces. h3manth.com/ai/webmcp/ #WebMCP #MCP #PromptAPI #WebAI #JavaScript

h3manth.com

WebMCP Bridge

Bridge any remote MCP server to Chrome's WebMCP API. Discover tools, prompts, and resources — then let browser AI agents use them.

001

Hemanth HM @gnumanth.bsky.social · 01/02/2026

As developers, we used to wait for standards. Now we just read the README. My latest post explores the "Cambrian Explosion" of AI Agent protocols (MCP, UCP, UTCP, A2A, ACP) and asks a hard question. h3manth.com/scribe/rfcs-... #Protocols #AI #Web

h3manth.com

RFCs vs. READMEs: The Evolution of Protocols | Scribe by Hemanth

How we build the internet's invisible contracts has changed completely. I'm not sure that's entirely good.

010

Hemanth HM @gnumanth.bsky.social · 21/01/2026

agentu now supports shorthand formats: user/repo, user/repo@branch. agent = Agent("assistant").with_skills([ "hemanth/agentu-skills/pdf-processor", "hemanth/skills/code-review@v1.0", # with branch/tag ]) pip install agentu pypi.org/project/agen... #python #agent #ai

pypi.org

agentu

A flexible Python package for creating AI agents with customizable tools

000

Hemanth HM @gnumanth.bsky.social · 12/01/2026

Baked an interactive demo for the Universal Commerce Protocol (UCP) UCP is an open standard enabling AI agents and platforms to complete purchases on any UCP-enabled merchant seamlessly. ucp-demo.web.app #UCP #AI

000