yo9esh

Yogesh Kumar Hi, I’m Yogesh, full-stack engineer.

I design how the pieces connect.

I build products end to end, and I care most about the parts in between: the pipelines, tunnels and protocols that decide whether a product actually works. Lately I go deep on a small set of tools instead of skimming ten stacks.

SDE intern at SystemStar Infotech · B.Tech CSE at Heritage Institute of Technology, ’28

See the systems Resume (PDF) 
$ curl -w "@timing" -so /dev/null yo9esh.tech
dns lookup
…
tcp connect
…
tls handshake
…
first byte
…
download
…
page ready
…
Real numbers from your browser’s Navigation Timing API. A 0 means the connection was reused.

Selected systems

07 case studies
01

Chatterbox TTS pipeline · 2026

From 48 GB accelerators to an 8 GB GPU

Automated voiceovers for 2,655+ real-estate listings. Listing text is cleaned into speakable text, then voiced by Chatterbox, whose memory grew with input length until it needed 48 GB accelerators.

  • 16–80 → 5 GBVRAM, any input length
  • −87%infra cost
  • 2,655+listings, zero OOM
Constraint
VRAM use was 16 GB at baseline and grew to 48–80 GB on long inputs, causing OOM crashes. Position overflows at token boundaries caused voice hallucinations.
Decision
Transformed S3Gen into bounded-memory chunked execution and patched the position overflows and token boundaries, so VRAM stays at ~5 GB at any length. Upstream, a Node.js sanitizer turns each listing into speakable text (house numbers, dimensions, currency, acronyms) capped at ~150 words.
PythonPyTorchCUDARunPodNode.jsMySQLBunnyCDN
02

SystemStar Infotech · 2025–26

A 360° player that starts fast

A 360° video web player and the video pipeline that feeds it, built at SystemStar.

  • 4K360° source footage
  • 2 sHLS segments (vs 4 s)
  • ABR720p → 4K ladder
Constraint
Generic CDN defaults use 4-second segments, which slows time-to-first-byte and playback start for heavy 4K footage.
Decision
FFmpeg probes each source, picks the fastest available encoder (GPU when present) and transcodes an adaptive ladder from 720p up to 4K into HLS with 2-second segments. A keyframe is forced every 2 seconds so every rendition switches on the same boundaries, and a master playlist feeds the React + Three.js player.
ReactThree.jsFFmpegHLSPythonCDN
03

CoralTeams · Coral hackathon · 2026

LAAF: infra from a potato laptop

CoralTeams is a self-hostable Teams interface I built for Coral, which connects tools like GitHub, Jira and Slack. It ships with a Docker image. The stack needed a 1 GB+ Rust engine and a 3 GB local LLM, and my laptop couldn’t run either.

  • $0per month
  • 4free platforms, 2 tunnels
  • 4 GB+stack, off-laptop
Constraint
No high-spec laptop, and cloud free tiers kept rejecting card verification. The stack couldn’t even be tested locally.
Decision
What I call CNG: the FastAPI backend runs in GitHub Codespaces (reached over SSH with the GitHub CLI), the LLM runs on Google Colab’s GPU, and ngrok tunnels expose both. The frontend sits on Vercel and the semantic engine with Coral on Render.
Tradeoff
Colab sessions expire and tunnel URLs rotate, so it’s for building and testing. Once everything works, it moves to a persistent server.
View 4 more systems Show fewer Cursor for n8n · AgentSFX · AlchEmaiLyst · Schedulo
04

1st prize · Ideatex, GDG on Campus HIT · 2025

Cursor for n8n

Describe an automation in plain English, get an n8n workflow. A Chrome extension built in five days with Soumyadeep Bhowmik.

  • Firstof 400+ developers
  • 5 daysidea to pitch
  • RAGprototype over n8n nodes
Constraint
LLMs hallucinate n8n nodes and parameters, so AI-generated workflows can’t be trusted.
Decision
Ground the model before it writes: the extension fetches the relevant n8n docs through Context7, adds them to the prompt for the chosen LLM, and returns workflow JSON ready to import. A RAG prototype went further, searching n8n’s node database and workflow templates (FAISS) with Gemini.
JavaScriptChrome extensionPythonRAGFAISSSQLiteGemini
Code
05

Track winner · Calcutta Hacks · Dec 2025

One-click game audio for Godot

Searching for and creating sound effects is a tedious part of making a game. We built a Godot add-on that creates SFX, character voices, NPC dialogue and BGM with AI and integrates them in one click. Team 4bajesonewale, with Soumyadeep Bhowmik, Harsh Sinha and Amlan Roy.

  • Track winnerCalcutta Hacks
  • 1 clickgenerate + integrate
  • 4person team
The AgentSFX dock in Godot listing generated sounds, next to the running platformer game
Constraint
Searching for and creating sound effects for a game is tedious.
Decision
An add-on inside the Godot editor: an LLM analyses the game’s code and a screenshot to list the sounds it needs, ElevenLabs generates them, and the add-on integrates everything into the code in one click.
06

Open source · GSSoC ’25 Project Admin

AlchEmaiLyst, and the community around it

An email assistant with three agents: Summarizer, Composer and Spam. Then I opened it up through GirlScript Summer of Code and ran it as a project.

  • 42★GitHub stars
  • 30+contributors onboarded
  • #1Alchemyst leaderboard
Constraint
Help people get through their inbox: summarise mail, draft and send replies, and clear spam from one screen.
Decision
A React app that signs in with Google OAuth 2.0, reads and sends mail through the Gmail API, and uses Alchemyst AI for summaries and drafts, falling back to Gemini.
07

Schedulo · Slack bot · Jan 2026

A Slack scheduler with local AI

Most companies won’t let internal data touch ChatGPT or Gemini. So I built a Slack bot that schedules meetings right inside Slack, with the AI running fully locally on Ollama.

  • LocalLlama 3.2 on Ollama
  • 0cloud LLM calls
  • 3 files+ config to set up
Constraint
Internal company data can’t be sent to cloud LLMs like ChatGPT or Gemini.
Decision
Slack talks to a Python agent over Socket Mode. A local Llama 3.2 on Ollama reads each mention and returns the intent, a summary and the start time as JSON. A Google Apps Script web app creates the Calendar event, and the bot confirms in Slack.
PythonOllamaSlackGoogle Apps Script
Code

Also built

  • Fullstack Odyssey Interactive simulators for the event loop, SQL joins, JWT rotation, Redis queues and Git’s DAG.
  • Lost & Found Campus lost-and-found with Gemini auto-tagging and a scored matching engine. Built in 2 hours; won.
  • Receipts.sh Telegram bot that logs promises in group chats, nudges before deadlines and deletes leaked API keys.
  • Moot Spiral Studio Photos + music → 60 FPS Fibonacci-spiral video, rendered entirely in the browser.
  • 3dWebImg Any 2D image into a 360° panorama, fully client-side.
  • ClimaVisor Climate forecasts and history on an interactive 3D globe.

Systems notebook

notes from the build
infra

LAAF: running a 4 GB AI stack from a potato laptop

A no-card, no-hassle way to run heavy services: Codespaces + Colab + ngrok, with the laptop as just the frontend.

Read → career

I shipped before the interview

Second year, tier-3 college, no referrals. The thing that got me an SDE internship had nothing to do with DSA.

Read →

On the podium, in the room

20+ hackathons
SIH 2024 Grand Finale: KrishiSahayak · Ministry of Fisheries, Animal Husbandry & Dairying
SIH 2024 Grand FinaleKrishiSahayak · Ministry of Fisheries, Animal Husbandry & Dairying
Ideatex, 1st prize: GDG on Campus HIT · Session Zero 2025
Ideatex, 1st prizeGDG on Campus HIT · Session Zero 2025
e-Summit ’25: two podiums: EDC, Heritage Institute of Technology
e-Summit ’25: two podiumsEDC, Heritage Institute of Technology
TCS hackathon win: A two-hour build, after 8+ tries with the same partner
TCS hackathon winA two-hour build, after 8+ tries with the same partner
Recognition
Jul 2026 Winner · TCS hackathon · Lost & Found, end to end in 2 hours
Dec 2025 Track winner · Calcutta Hacks · AgentSFX, AI game audio for Godot
Nov 2025 1st runner-up ×2 · e-Summit ’25, EDC HITK · Ideathon (Optiq) + Marketing Gimmick
Jul 2025 1st prize · Ideatex, GDG on Campus HIT · Cursor for n8n · 400+ developers
Feb 2025 2nd runner-up · ACM Brainstorm quiz
Dec 2024 Grand Finalist · Smart India Hackathon · Top 0.2% of 50k+ teams · team KrishiSahayak
Sep 2024 3rd of 200 · HackHeritage · SIH college prelims

Few tools, all the way down

depth over breadth
ssh
SSH into GitHub Codespaces with the GitHub CLI to run heavy services remotely (LAAF).
ngrok / tunnels
Joined three free clouds into one dev stack for LAAF.
FFmpeg + HLS
The 4K → adaptive-stream pipeline at SystemStar.
TypeScript · Node
The listing-text sanitizer behind the TTS pipeline; AlchEmaiLyst’s OAuth and Gmail flows.
Python · PyTorch · CUDA
Rewrote model inference to fit an 8 GB GPU.
MySQL · SQLite
MySQL in the TTS pipeline; the SQLite node knowledge base behind Cursor for n8n.
React · Three.js
The 360° player at SystemStar, ClimaVisor, 3dWebImg.

Right now: going deep on ssh, curl, ngrok and system design.

Yogesh in a yellow kurta at an old Kolkata gate, smiling

About

From Gaya to Kolkata, one shipped thing at a time.

I’m a Computer Science undergrad at Heritage Institute of Technology, Kolkata, and a software engineering intern at SystemStar Infotech. I like the unglamorous middle of a product: how 4K 360° video starts playing without the wait, how an LLM gets grounded in real data, how a stack runs when the hardware says it can’t.

How I work: I show work early. I got my internship by building a mock before the interview, not after. Here’s that story.

I’ve been through 20+ hackathons, from a two-hour build to the SIH Grand Finale, and I ran AlchEmaiLyst as an open-source project in GSSoC ’25.

Contact

Have a system that needs its pieces connected?

fbyogesh111@gmail.com