Om Patnaik — AI and robotics portfolio

CLASS XII · DELHI · TARGETING AY 2027–28

Building AI
from first
principles.

From transformer internals to physical reasoning robots — I build the thing, then ask what's underneath it. Currently working through the TARQ curriculum: 10 projects, 14 weekends, one reasoning robot.

GITHUB ↗

// ABOUT

I'm Om Patnaik, a Class 12 student in Delhi working at the intersection of machine learning, robotics, and hardware. My first real research project — comparing mBERT and MuRIL on Hinglish hate speech — taught me to read what a model's failure is actually saying.

That question led deeper: from language, to reasoning systems, to physical hardware. I'm now building TARQ — a 6-DOF arm that reasons before it acts. The goal isn't a demo. It's understanding every layer of the stack, from training data to silicon.

// PROJECTS

002 · CAPSTONE IN PROGRESS
TARQ — Physical Reasoning Robot
A 6-DOF servo arm on Raspberry Pi 5. Listens, sees, reasons before acting. Uses Moondream VLM + Llama-3. Asks when uncertain. Refuses unsafe commands — enforced at both prompt and code level.
Raspberry Pi 5Llama-3MoondreamPyTorchFastAPIOllama
— coming soon
003 · HARDWARE IN PROGRESS
FPGA Neural Network Accelerator
AG News classifier compiled to Verilog via a Python script. Synthesised on Basys3 Artix-7. Vivado utilisation: Block RAM = 0. Weights aren't fetched from memory — they are the circuit.
VerilogVivadoBasys3 FPGAINT8scikit-learn
— coming soon
004 IN PROGRESS
Transformer from Scratch
Self-attention, multi-head attention, positional encoding — every component in PyTorch, no high-level abstractions. Character-level language model trained on Shakespeare.
PyTorchPython
— coming soon
005 UPCOMING
Deployed ML API
Text classifier wrapped in FastAPI, containerised with Docker, deployed to Railway and AWS Lightsail with nginx.
FastAPIDockerAWSnginx
— coming soon
006 UPCOMING
RL Robot Arm in Simulation
Kuka iiwa arm trained to reach targets in PyBullet using PPO and SAC. Reward shaping and curriculum learning before any hardware.
PyBulletPPO/SACStable-Baselines3
— coming soon

// RESEARCH

mBERT vs MuRIL — Hinglish Hate Speech Detection

Compared two multilingual transformer models on Hinglish hate speech detection — mBERT (Google's multilingual BERT) against MuRIL, a model pre-trained specifically on Indian languages. The core question was not which number was higher, but what mBERT's failure mode reveals about how multilingual models represent code-switched text.

FINDING → Code-switching breaks assumptions baked into clean monolingual pre-training data. mBERT's failure on Hinglish exposes gaps that corpus statistics alone don't predict — MuRIL's Indian-language focus handles the distribution shift more gracefully.

// SKILLS

LANGUAGES

  • Python 3.11
  • Verilog
  • Bash

ML / AI

  • PyTorch
  • scikit-learn
  • Ollama
  • Stable-Baselines3
  • OpenCV

INFRA

  • Docker
  • FastAPI
  • Git
  • AWS Lightsail
  • Railway

HARDWARE

  • Raspberry Pi 5
  • Basys3 FPGA
  • Vivado
  • PCA9685 / I2C
  • IKPy

MODELS

  • Llama-3 (8B)
  • Moondream 1.7B
  • YOLOv8
  • ResNet-18
  • mBERT / MuRIL

greyed out = in progress or upcoming as part of the TARQ curriculum

// CONTACT

If you're from an admissions office, a research group, or just curious about something I've built — reach out.

om@yourdomain.com GitHub ↗