OMY_AIML_SPPU_2027

I BUILD THE
BACKEND HALF
OF AI PRODUCTS.

OM YADAO AI/ML ENGINEER PUNE, MAHARASHTRA, IN FINAL YEAR B.E. AI & DS

Inference APIs, retrieval pipelines and agent systems that hold up outside a notebook. Three AI/ML internships, eight-plus shipped projects, and a preference for the parts of the stack nobody demos.

Selected work* Résumé ↓ Contact

Scroll to index

0AI/ML internships

0Shipped projects

0Deepfake classifier accuracy

0B.E. AI & Data Science

00 / INDEX

Seven sections. Everything here is a system that ran, a role that happened, or a course that finished.

01AboutProfile / working on 02Selected work05 systems + 03 builds 03Experience03 internships 04Stack08 categories 05EducationSPPU / Maharashtra board 06ContactOpen to roles

01 / PROFILE

About

Final-year AI & Data Science student at SPPU, Pune. Backend-first.

I'm a final-year Artificial Intelligence and Data Science student at SPPU, Pune. Most of my work sits behind the model rather than on top of it: serving inference over HTTP, wiring retrieval, keeping pipelines from falling over when a provider times out.

I care about how these systems actually work internally, not just which library call returns the right shape. That's why a lot of what I build runs locally — quantised models under Ollama, embeddings and vector search on my own machine — before it ever touches a hosted API.

Right now I'm going deeper on PyTorch and parameter-efficient fine-tuning, and building domain assistants where being wrong has a cost.

Fine-tuning with LoRA and QLoRA~~IN PROGRESS~~

Model quantisation and local serving~~IN PROGRESS~~

Agent orchestration with LangGraph~~IN PROGRESS~~

Production deployment with BentoML~~IN PROGRESS~~

PROFILE — QUICK FACTS

RoleAI/ML Engineer

FocusBackend · Inference · Retrieval

Based inPune, Maharashtra, IN

EducationB.E. AI & DS — SPPU, 2027

StatusOpen to roles & internships

02 / SELECTED WORK

Projects

Five systems, plus the smaller builds underneath them.

PROJECT 01FEATUREDORCHESTRATION

MAP — Multi-Agent AI Automation Platform

Architecture and review lead, team of four

A production-grade orchestration platform where independent agents pick up work, call tools and hand results down a pipeline. I designed the architecture and reviewed every pull request into main.

The interesting problems were the boring ones: what happens when one agent stalls, how a failed step retries without duplicating side effects, and how you trace a single request across a distributed worker pool. Circuit breakers, structured logging and idempotent task dispatch answer most of it.

FastAPILangGraphPostgreSQLRedisCelerystructlogReact 18TypeScript

DIAGRAM — REQUEST LIFECYCLEMAP / V1

SRSCLASSIFICATION

Support Resolution System

Multi-tenant ticket classification backend

An enterprise support system that reads an incoming ticket and routes it to the right queue at 90%+ accuracy, isolated per tenant so one customer's volume can't starve another's.

Classification runs against OpenAI first and falls back to a self-hosted BentoML model when the API is slow, rate-limited or down, so routing degrades instead of stopping. Redis caches repeat classifications and slowapi caps per-tenant throughput.

TICKETREDIS CACHEOPENAIBENTOML FALLBACKQUEUE

Routing90%+ accuracy

IsolationMulti-tenant

FallbackOpenAI → BentoML

Limitsslowapi per tenant

FastAPIPostgreSQLRedisBentoMLslowapiOpenAI API

SHAWTYAGENT

Local-First Coding Agent

Autonomous agent, zero cloud dependency

A Claude Code-style agent that reads a repository, plans an edit and applies it — running entirely against local models under Ollama, so no source code leaves the machine.

Built on a LangGraph state machine with tool nodes for file reads, writes and shell commands, and an explicit review step before anything is written to disk.

REPOLANGGRAPH STATE MACHINEREVIEWWRITE

RuntimeFully local

ControlLangGraph state machine

ToolsRead / write / shell

GuardReview before write

PythonLangGraphOllamaQwen2.5 Coder

DEEPFAKECOMPUTER VISION

Deepfake Detection API

Computer vision service with explainability

An EfficientNet-B4 classifier that flags manipulated faces at 94%+ accuracy, served behind a Dockerised API with a React frontend for real-time inference.

Every prediction returns a GradCAM heatmap alongside the score, so a reviewer can see which region of the face drove the call instead of trusting a bare number.

FACEEFFICIENTNET-B4GRADCAMSCORE + HEATMAP

Accuracy94%+

BackboneEfficientNet-B4

ExplainabilityGradCAM heatmap

DeliveryDocker + React

PyTorchEfficientNet-B4GradCAMFastAPIDockerReactTypeScript

SKINCAREFINE-TUNING

AI Skincare Assistant

Fine-tuned Qwen on curated medical data

A fine-tuned assistant that gives skincare routines grounded in medical literature rather than marketing copy, with product suggestions priced for the Indian market.

Training data was assembled from PubMed abstracts, MedQuAD and a Kaggle skincare set into roughly 5,300 curated examples, then used to fine-tune Qwen with PEFT. Most of the effort went into cleaning the dataset, not the training run.

PUBMEDMEDQUAD5,300 EXAMPLESPEFT FINE-TUNE

Dataset~5,300 examples

SourcesPubMed / MedQuAD / Kaggle

MethodPEFT on Qwen

MarketIndia-priced products

PythonHugging FacePEFTQwenLoRA

ALSO BUILT

Local RAG pipeline

Chunking, local embeddings and vector search with Ollama, FAISS and ChromaDB — EmbeddingGemma and Qwen2.5 Coder 1.5B, fully offline.

Human action recognition

Real-time webcam pose detection and action classification with MediaPipe, OpenCV and Keras.

Image captioning

CNN encoder with an LSTM decoder trained on Flickr8k to generate captions from scratch.

03 / TRACK RECORD

Experience

Three AI/ML internships, all backend or model-side.

Kemuri
Technology

Dec 2025 — Present

Remote / AI/ML Intern

AI/ML Intern, backend. Working on the Clean Ocean Ensemble waste-classification system and the Python services around it.

01Prompt engineering for vision-language models against real field photography
02OCR prompt optimisation and a weight-estimation pipeline for scale readings
03FastAPI endpoints, Pydantic schemas and image preprocessing utilities
04Pytest coverage, security hardening and performance work on existing services
05Remote development on a Mac Mini over VS Code Remote SSH, feature-branch workflow

Evoastra
Ventures

Oct 2025 — Nov 2025

AI/ML Intern

AI/ML Intern on predictive analytics and applied ML work.

01Model development and data preprocessing for live production projects
02Performance tuning on existing training and inference code

Future
Interns

Data Science Intern

Programme

Structured data science programme covering end-to-end machine learning workflows and exploratory analysis.

04 / TOOLKIT

Stack

What I reach for. No proficiency bars — everything listed has shipped in something above.

Languages

Python, C, C++, SQL, Bash

Backend

FastAPI, Celery, Docker, Pydantic, structlog, BentoML, REST

Machine learning

PyTorch, TensorFlow, Keras, OpenCV, MediaPipe, scikit-learn

LLMs and agents

LangGraph, LangChain, RAG, Ollama, Hugging Face Transformers, PEFT, LoRA, QLoRA, OpenAI API

Data and retrieval

PostgreSQL, Redis, MySQL, FAISS, ChromaDB, NumPy, Pandas

Frontend

React 18, TypeScript, HTML, CSS

Environment

Ubuntu, Git, tmux, SSH, NVIDIA CUDA setup, Tailscale, Jupyter, Colab

Practice

Feature branches, pull-request review, Pytest, test-driven development

CURRENTLY LEARNING

PyTorch internals, LoRA and QLoRA fine-tuning, model quantisation, BentoML deployment and agent development with LangGraph.

05 / BACKGROUND

Education

Currently in the final year.

2023—2027

B.E. Artificial Intelligence & Data Science

Sinhgad Institutes (SRCOE), Savitribai Phule Pune University — final year

2021—2023

Higher Secondary, Maharashtra Board

SBES College of Science, Chh. Sambhajinagar

2020—2021

Secondary, Maharashtra Board — 84.40%

Godavari Public School, Chh. Sambhajinagar

LANGUAGES EnglishHindiMarathiJapanese (spoken)

06 / CONTACT

Open to AI/ML engineering roles and internships.

If you're building something that needs models served, retrieved over, or wired into a real backend, I'd like to hear about it.

Email me* Résumé ↓

DIRECT LINES

Emailomyadao1706@gmail.com↗ Phone+91 89994 62510↗ GitHubgithub.com/Yad4o↗ LinkedInlinkedin.com/in/om-yadao-5359791a9↗ RésuméOm-Yadao-Resume.pdf↓

I BUILD THE BACKEND HALF OF AI PRODUCTS.

Contents