Jayant Kothari — AI/ML Engineer

About

Background

I'm a Computer Science undergraduate at IIIT Kota and an AI/ML engineer focused on taking models out of notebooks and into real, scalable products.

My work spans applied machine learning end to end — designing hybrid-retrieval RAG pipelines, fusing supervised and unsupervised models for fraud detection, and deploying inference as low-latency FastAPI microservices with Docker.

I care about the details that make systems trustworthy: explainability with SHAP, grounded LLM responses, caching that cuts cost, and metrics that hold under p95 latency budgets.

IIIT Kota

B.Tech, Computer Science & Engineering

Duration2023 – 2027

GPA8.15 / 10

LocationKota, India

Experience

Where I've worked

Jan 2026 – Mar 2026 Remote

Machine Learning Intern

Gyanama

Developed an end-to-end analytics platform using XGBoost to predict academic performance, attendance patterns, and institutional health scores across 15+ partner schools.
Constructed feature pipelines over 10,000+ records with Pandas & Scikit-learn, achieving 87% accuracy and 0.91 ROC-AUC.
Designed automated evaluation dashboards — cutting manual analysis effort by 60% and report turnaround from 3 days to 4 hours.
Deployed model inference as FastAPI microservices, serving real-time predictions at p95 latency under 200ms.

Projects

Selected work

01

DocuMind

Multi-PDF RAG System

Architected an asynchronous multi-stage RAG pipeline (FastAPI + Celery + LangGraph) with 4-level hierarchical chunking and hybrid retrieval (FAISS + BM25 + RRF), handling 1,000+ page PDFs with zero memory spikes.
Implemented 5 LangGraph reasoning workflows with Redis TTL caching, cutting redundant LLM calls by 40% and query latency by 55%.
Containerized 5 services via Docker Compose with health checks and SSE streaming delivering tokens in under 800ms.

PythonFastAPILangGraphFAISSCeleryRedisDocker

Live Demo View Code

02

Credit Card Fraud Detection

Real-time Fraud Decisioning

Built a real-time microservice fusing a supervised XGBoost classifier (PR-AUC 0.88, ROC-AUC 0.98) with an unsupervised deep autoencoder to catch known and zero-day fraud across 284,807 transactions.
Designed a 3-tier graded decision engine (APPROVE / VERIFY OTP / BLOCK) with a simulated OTP challenge for medium-risk cases.
Integrated SHAP attribution with a generative-AI layer (Gemini 2.5 Flash-Lite) to explain every decision in plain English, returning predictions in ~1.25s.

PythonXGBoostTensorFlowSHAPFastAPIStreamlitDocker

Live Demo View Code

03

YouTube Smart Chatbot

Multilingual Conversational RAG

Built a multilingual conversational RAG system over YouTube transcripts with timestamp-synced captions, supporting videos up to 12+ hours via parallel 300s Deepgram Nova-2 chunks.
Engineered a dual-source ingestion flow (native captions + speech-to-text fallback) preserving word-level timing across 20+ languages, improving coverage by 35%.
Tuned prompt templates (temp 0.3–0.4) for grounded responses, reaching sub-3-second answers via Groq with 10-turn stateful memory.

PythonStreamlitFastAPIGroq Llama 3.3Deepgramyt-dlp

View Code

Skills

Tools & technologies

Languages

Python, C++, SQL, Java

AI / ML

Generative AI, NLP, LLM Fine-tuning, RAG, Prompt Engineering, Deep Learning, XGBoost

Frameworks

LangChain, LangGraph, Hugging Face, Sentence-Transformers, PyTorch, Scikit-learn, Pandas, NumPy

Retrieval

FAISS, BM25, Cross-Encoder Re-ranking, Hybrid Search, Embeddings

Backend & Data

FastAPI, Streamlit, Gradio, REST APIs, Celery, Redis, PostgreSQL, SQLite, SQLAlchemy

DevOps & Tools

Git, GitHub, Docker, Jupyter, Google Colab

Achievements

Beyond the code

600+

DSA problems solved on LeetCode ↗

3★

Rating on CodeChef ↗