LLM Pretraining AI Infra & Systems 3× Startups Building Vithos ↗ Teaching Hard ML

Aditya Ghai

I work on hard problems — pretraining large language models, designing AI infrastructure, and shipping production-grade ML at 3 startups. When I'm not building, I break down the hard parts of ML on YouTube.

Startups Shipped At
Hackathon Wins
225M Param LLM from Scratch
800K Users Served
Vithos AI Health Platform
01

Experience

Building and shipping ML in production across 3 startups — from LLM systems to on-device computer vision.

Sept 2025 — Present

SDE Intern (ML + Dev)

Rimo LLC, Japan Remote

  • Architected and shipped RAG chat end-to-end — a production meeting transcript search system serving 800K users — on a 3-service microarchitecture (Go/Python + Cloud Run Jobs), with full ownership from design to deployment monitoring.
  • Built a real-time SSE streaming pipeline combining BM25 + Dense vector search (Ruri embeddings), RRF, and RuriV3 reranking; reduced query latency by 12% handling ~6,500 req/sec in production.
  • Designed an async distributed indexing system on GCP Cloud Run Jobs (NVIDIA L4 GPU) for chunking, embedding generation, and MongoDB Atlas writes — decoupled from the API layer to eliminate resource contention at scale.
Jan 2025 — Sept 2025

AI Engineering Intern

3S SmartShips, Singapore Remote

  • Built a custom Gauge OCR pipeline for shipboard automation — collected and augmented domain-specific datasets, trained RF-DETR and YOLOv11 models optimized for mobile offline inference, and developed a custom algorithm for accurate gauge reading.
  • Deployed across 750+ dock ships via an Android app, fully automating repetitive manual inspection workflows for Tata NYK shipping operations.
02

Teaching Hard ML

I break down the parts of ML most tutorials skip — pretraining, distributed training, and real system design. No wrappers, no hand-waving.

03

Projects

04

Achievements

Convolve 4.0 Winners
Feb 2026

Convolve 4.0 — 1st Runner Up

IIT Guwahati Pan IIT Competition

Built an end-to-end local OCR pipeline for handwritten & vernacular texts under strict VRAM and latency constraints, achieving 82% DLA. Built for IDFC First Bank.

Kaggle Notebooks Expert
Jun 2025

Kaggle Notebooks Expert

Kaggle

Recognized as a Notebooks Expert on the Kaggle platform, though I am not very active here :)

MLH HackByte 2.0 Winner
Apr 2024

MLH HackByte 2.0 Winner

IIITDM Jabalpur

Developed a deep learning model that takes badminton videos and provides overall player analysis — detailed reports and graphs on shot selection and movement.

IIT Indore AnalyticaX
Mar 2024

AnalyticaX ML Competition Winner

IIT Indore

Complete EDA + modelling on a complex real-world vaccination dataset, achieving 0.872 ROC-AUC with stratified cross-validation.

ThermaOracle IIT Roorkee
Mar 2024

ThermaOracle ML Competition Winners

IIT Roorkee

Surpassed teams from top IITs in an advanced forecasting & regressive temperature prediction competition with intricate data design.

IIT Jodhpur BioInnovate
Jan 2024

AI BioInnovate Hackathon Champion

IIT Jodhpur

Devised a computationally efficient clustering approach for the Moses dataset — 1.5M molecules in SMILES format — maintaining high accuracy while minimizing resources.

05

Writing

Latest articles from Medium.

Loading articles…
06

Let's Connect

Aditya Ghai

Feel free to reach out for collaborations, research discussions, or just a friendly chat about ML & AI. I'm currently building Vithos and teaching hard ML on YouTube.