AI Engineer

Chaitanya
Kumar

Building end-to-end intelligent systems

I design and deploy production AI systems — from LLM pipelines and RAG architectures to full-stack ML applications. I don't just use APIs; I understand what happens inside them.

Things I've Built

AI Data Analyst Agent

Automates end-to-end ML pipelines for non-technical users — no code required.

A Flask web app that ingests any CSV, auto-detects classification vs. regression, engineers features, trains Random Forest models, and delivers a professional PDF report. The tool-calling chatbot (GPT-4o-mini) operates directly on live session data — eliminating hallucinated metrics entirely.

// System Architecture
CSV Upload
Auto Cleaning + EDA
Feature Engineering
Random Forest (sklearn)
GPT-4o-mini Tool Calls
PDF Report + Chat
Tech Stack
Python Flask OpenAI GPT-4o-mini scikit-learn Pandas ReportLab Gunicorn Render
Key Metrics
0
LLM Hallucinations
Live
Deployed on Render
Auto
Task Detection

Vendor Analytics Dashboard

Full-stack vendor registration + real-time analytics platform on serverless cloud.

End-to-end web system for the Unified Rewards System (URS) with vendor registration, secure Firebase auth, and live sales analytics dashboards. Features 5-minute auto-refresh, date-range filtering, and one-click CSV/PDF export.

// System Architecture
Vendor Registration
Flask REST API
Firebase Firestore
Chart.js Dashboards
CSV / PDF Export
Tech Stack
Python Flask Firebase Firestore JavaScript TailwindCSS Chart.js Render
Key Features
5m
Auto Refresh
HTTPS
Production
NoSQL
Serverless DB

Multimodal Image Sentiment Analysis

Fuses vision + language to classify image sentiment — going beyond pixels.

A deep learning fusion pipeline combining ResNet50 visual features (2048-dim) with BERT-encoded text extracted via Tesseract OCR (768-dim) into a joint dense classifier for Positive / Negative / Neutral sentiment.

// Fusion Architecture
Image Input
ResNet50 (2048-dim)
Feature Concat
Sentiment Class
OCR (Tesseract)
BERT (768-dim)
Feature Concat
Tech Stack
Python TensorFlow BERT ResNet50 Tesseract OCR HuggingFace
Key Metrics
86.8%
Train Accuracy
10
Epochs
bs=32
Batch Size

E-Commerce Web Application

Full-stack commerce platform with MVC architecture and responsive design.

A production-grade e-commerce app built with Flask featuring product listings, cart management, and SQLite-backed data layer — fully decoupled routing, data access, and Jinja2 templates.

// MVC Architecture
User Request
Flask Router
database.py (CRUD)
SQLite
Jinja2 Template
Tech Stack
Python Flask SQLite HTML/CSS JavaScript Jinja2
Key Features
MVC Pattern CRUD Operations Responsive UI Shopping Cart

The Stack

🧠
AI / ML Core
LLMs & RAG
Transformers / BERT
scikit-learn / XGBoost
TensorFlow / PyTorch
Computer Vision (CV)
⚙️
Backend & APIs
Flask / FastAPI
Python
REST API Design
SQL (MySQL / SQLite)
Firebase / Firestore
🚀
Deployment & Cloud
Render / Vercel
AWS (EC2, S3, Lambda)
Docker
AWS SageMaker
Kubernetes
🔗
LLM Ecosystem
OpenAI API
LangChain
HuggingFace
LLaMA / Open Models
Prompt Engineering

How I Think

01
How I design AI systems
  • Start with problem framing — what's the actual failure mode being solved?
  • Decide ML vs. rule-based: if edge cases are enumerable, rules win
  • Build the pipeline first, then optimize the model inside it
  • Design for degradation — what happens when the model is wrong?
  • Ship a deterministic baseline before adding probabilistic layers
02
How I handle LLM hallucination
  • Tool-calling over direct generation: compute facts, don't describe them
  • Ground responses in live data sources — retrieval before generation
  • Validation layers: parse structured output, check ranges, fallback gracefully
  • Separate "thinking" from "output" — structured reasoning prompts
  • Test with adversarial inputs, not just happy paths
03
How I optimize inference performance
  • Profile first — latency is almost never where you think it is
  • GPU batching: queue multiple requests, amortize model load time
  • Cache embeddings for static content — avoid recomputing
  • Model quantization before scaling hardware (cheaper wins first)
  • Monitor cold-start vs. warm latency separately in production
04
Flask vs. FastAPI — when and why
  • Flask for rapid prototypes, AI demo apps, Jinja2 server-side rendering
  • FastAPI for async ML inference endpoints, type-safe APIs, high throughput
  • Flask's simplicity avoids overhead when concurrency isn't the bottleneck
  • FastAPI's Pydantic models enforce schema contracts — critical in production
  • The right choice depends on the team, not just the benchmark

Where I've Worked

Jul 2025 – Nov 2025
Business Analyst
ICICI Bank
Business Analyst — Corporate Healthcare
  • Managed a portfolio of 10+ corporate healthcare clients worth ~$500K+, analyzing financial statements and operational data.
  • Conducted market and client analysis to identify cross-selling opportunities, directly supporting revenue growth strategies.
  • Structured client portfolio data to derive profitability insights, improving prioritization and decision-making efficiency.
Jan 2025 – Feb 2025
AI Intern
GenScriptAI
AI Intern — Multimodal Generation
  • Built end-to-end AI pipelines using LLM and TTS models to generate audio from text, improving coherence via prompt engineering.
  • Developed a multi-modal system combining LLMs, diffusion models, and audio synthesis to automate video creation from text.
  • Optimized image generation with GPU acceleration — reduced inference time from ~5 minutes to ~15 seconds (20× speedup) via batching.
  • Deployed Flask-based APIs to manage generation workflows, enabling scalable multi-pipeline execution.
May 2024 – Jul 2024
Tech Intern
Jindal Steel Limited
Technology Intern — ML & Optimization
  • Worked with large-scale operational data — data cleaning, feature engineering, and EDA to identify key performance drivers.
  • Built an ML model achieving R² ≈ 0.85, significantly reducing system output variability.
  • Developed a data-driven efficiency tool identifying 10–15% optimization opportunities through input-output analysis.
  • Analyzed time-series trends to surface patterns affecting system efficiency, supporting data-driven decisions.

Wins & Awards

🏆
NPCI Fintech Hackathon Finalist
National Payment Corporation of India
🥇
IIT Kharagpur AI-volution — Top 8
Out of 800+ teams nationwide
🎖️
Director's Awardee
Best Performer — Content Team Lead, Counseling Cell
🌐
Inbound Coordinator
International Relations Cell, IIT Hyderabad
CK
IIT Hyderabad · 8.54 GPA

Building AI that
actually works in production

I'm a B.Tech graduate from IIT Hyderabad (2021–2025) with a deep focus on end-to-end AI engineering. I'm not a researcher — I'm a builder. I care about the gap between a model that works in a notebook and a system that works for real users.

My experience spans LLM applications, multimodal AI, production ML systems, and full-stack development. I've deployed inference pipelines at GenScriptAI, built ML models for industrial optimization at Jindal Steel, and shipped multiple live web applications.

I'm targeting AI Engineer roles where I can own the full stack — from model selection to production deployment — at companies building with AI as a core product, not an afterthought.

LLM Systems ML Pipelines API Engineering Cloud Deployment Multimodal AI RAG Architecture

Let's build
something real

I'm actively looking for AI Engineer roles. If you're building intelligent systems and need someone who cares as much about architecture as about accuracy — let's talk.