"A deep dive into my journey as an AI engineer, featuring multilingual voice assistants, teaching tools for India, and personalized AI systems. Published researcher in IEEE OTCON 2025."
Inside an AI Engineer's Portfolio: Building Solutions That Actually Matter
Hey there! I'm Aadya Madankar, a Generative AI & Machine Learning Specialist from Nagpur, India. I graduated from Priyadarshini Engineering College, Higna Road, and I believe that great code doesn't just execute commands—it learns, adapts, and creates solutions.
You know what? The world of AI engineering can feel overwhelming with its constant barrage of new frameworks, models, and hype. But here's what I've learned: the best projects aren't the ones using the shiniest tech—they're the ones solving real problems for real people.
I'm a college graduate with a strong foundation in data science, specializing in machine learning and deep learning. My experience includes active participation in Kaggle competitions and collaborative GitHub projects, demonstrating proficiency in OpenCV-Computer Vision and Generative AI LLM models.
Let me walk you through my portfolio and share what building production-ready AI systems has taught me.
🎯 The Philosophy: Impact Over Impressiveness
Before diving into the projects, here's my core belief: The world is one big data problem.
Every project in my portfolio stems from identifying a genuine gap where AI can make a measurable difference. Not "AI for AI's sake," but intelligent systems that address accessibility, education, and productivity challenges.
My Specializations:
- Generative AI & Large Language Models (LLMs): Building conversational agents, multimodal systems, and intelligent assistants
- Machine Learning & Deep Learning: From computer vision to predictive modeling
- Data Science: OpenCV-Computer Vision, NLP, and data-driven decision making
- Deployment & Production: Taking models from Jupyter notebooks to real-world applications
You can check out my full portfolio at aadyamadankar.life, but let me break down the work that taught me the most.
🗣️ AI-Associate: Breaking Language Barriers with Voice AI
The Problem
India has 22 officially recognized languages and hundreds of dialects. Yet most voice assistants only work well in English and maybe Hindi. Millions of people are locked out of voice technology simply because they speak Marathi, Tamil, Telugu, or any other regional language.
What I Built
A production-ready voice assistant supporting 30+ Indian languages with real-time multimodal processing. This isn't a wrapper around existing APIs—it's an intelligent routing system that handles:
- Culturally aware responses (understanding context matters more than literal translation)
- Multimodal processing (text, voice, and visual inputs)
- Real-time inference with optimized latency for practical use
The Tech Stack
- Speech Recognition: Custom ASR models fine-tuned for Indian accents
- LLM Integration: Google Gemini for multilingual understanding
- Deployment: Vercel for edge-optimized serving
- Monitoring: Real-time performance tracking across languages
What I Learned
This project taught me that accessibility isn't just a feature—it's a design constraint. When you're building for linguistic diversity, you can't just translate; you need to understand cultural context, regional idioms, and varying levels of digital literacy.
Published my findings in IEEE OTCON 2025 (4th OPJU International Technology Conference on Smart Computing for Innovation and Advancement in Industry 5.0) in a research paper titled "AI-Associate: A Lightweight Architecture for Conversational Agents" co-authored with U.A.S. Gani, Atharv Shinde, Atharva Sonwane, and team. The paper demonstrates how scalable architecture can enable culturally inclusive conversational AI.
👩🏫 Saahayak: AI Teaching Assistant for Rural India
The Reality Check
Picture this: one teacher managing three different grade levels in a single classroom with minimal resources. This is the reality in many rural Indian schools.
Teachers spend hours creating differentiated worksheets, visual aids, and lesson plans—time they could spend actually teaching.
The Solution
Saahayak (Sanskrit for "helper") is an AI-powered teaching assistant that generates:
- Hyper-localized educational content
- Differentiated worksheets for multi-grade classrooms
- Visual aids and lesson plans
- All from text, voice, or image inputs in 25+ Indian languages
Why This Matters
Built with Google Gemini and Genkit, this project demonstrates how practical AI can save educators hours of preparation time while maintaining the human touch that makes great teaching possible.
It's not about replacing teachers—it's about giving them superpowers.
Technical Highlights
- Multimodal input processing: Upload a textbook page photo, get lesson plans
- Language flexibility: Works seamlessly across Hindi, Marathi, Telugu, and more
- Offline-first design: Considering limited internet connectivity in rural areas
- Context-aware generation: Understands the Indian curriculum framework
🧬 Custom SLM: Training My AI Clone
This one's my favorite because it's a bit meta.
The Concept
I'm training a Small Language Model on my experiences, knowledge, and problem-solving patterns—essentially creating an AI version of how I think, code, and approach challenges.
Why Build This?
- Knowledge preservation: Capture my expertise in a queryable format
- Scalable mentorship: Help others even when I'm not available
- Living portfolio: Demonstrates both technical capability and philosophical understanding
- Learning tool: Understanding how to distill personal expertise into training data
The Process
This isn't just fine-tuning a model on my GitHub repos. It involves:
- Data collection: Code, documentation, problem-solving approaches, design decisions
- Pattern extraction: Identifying recurring themes in how I approach problems
- Continuous learning: The model evolves as I do
- Ethical boundaries: Being transparent about what it is (and isn't)
The Philosophy
This project bridges AI engineering with self-documentation. It's not about replacing myself—it's about creating an accessible interface to my knowledge and demonstrating how SLMs can be personalized tools, not just generic assistants.
🛠️ The Tech Stack: Tools I Actually Use
Here's what's in my daily toolkit (and why):
Core ML/AI
- TensorFlow & Keras: For custom model training
- PyTorch: When I need more flexibility
- LangChain: RAG systems and agent orchestration
- Hugging Face: Model experimentation and deployment
API & Deployment
- FastAPI: Lightning-fast API development
- Streamlit: Rapid prototyping and demos
- Docker: Containerization for reproducible deployments
- Vercel: Frontend hosting with edge optimization
Data & Databases
- Pandas & NumPy: Data manipulation foundation
- MongoDB: Document storage for unstructured data
- ChromaDB & Faiss: Vector databases for RAG systems
Development Workflow
- Git/GitHub: Version control and collaboration
- Jupyter: Experimentation and documentation
- VS Code: Primary IDE with AI extensions
- Kaggle: Dataset exploration and competitions
📚 Other Projects Worth Mentioning
Beyond the flagship projects, here are some other notable works that showcase different aspects of my AI engineering skills:
Multimodal PDF Assistant
RAG-based system for answering queries from PDFs using both text and images. Think "ChatGPT for your research papers" but with vision capabilities. Perfect for students and researchers who need to quickly extract insights from dense academic papers.
Tech: LangChain, Google Gemini Pro, Streamlit, RAG, FAISS, PyPDF2
Voice-to-Image Generator
Generate images from voice input in under a second using NVIDIA TensorRT optimization. Just speak what you want to see, and the system creates it in real-time. A fascinating exploration of multimodal AI that combines speech recognition with high-speed image generation.
Tech: SDXL Turbo, NVIDIA TensorRT, Stable Diffusion XL, ASR, CLIP, U-Net, VAE
Multi-Modal Screen Assistant
AI-powered desktop assistant combining visual processing, text analysis, and voice interaction. It's like having a programming companion that can see your screen, understand your code, and help debug or suggest improvements through natural conversation.
Tech: OpenAI Whisper, Google Gemini, Groq, PyAudio, Pillow
RAG Notebooks Repository
Comprehensive collection of advanced RAG (Retrieval-Augmented Generation) techniques—my knowledge base for building production-ready retrieval systems. This repository serves as both a learning resource and a practical guide for implementing state-of-the-art RAG approaches.
Tech: LlamaIndex, VectorStores, OpenAI, Gemini, Cohere, Hugging Face, ChromaDB
Advance-RAG-with-Langchain
Deep exploration of advanced chatbot techniques using LangChain. Covers everything from basic conversational AI to complex multi-agent systems with web search capabilities, database integration, and custom tool usage.
Tech: OpenAI, Groq, Streamlit, LangChain, LangServe, BeautifulSoup, ChromaDB, Wikipedia API
Crew-AI Multi-Agent System
Python-based multi-agent AI system built with CrewAI framework. Demonstrates how autonomous agents can collaborate to solve complex tasks that would be difficult for a single agent to handle alone.
Tech: CrewAI, Hugging Face, Python
NVIDIA Model Deployment with LangServe
Deploy NVIDIA's GPU-accelerated AI models as APIs using LangServe. Shows how to take advantage of NVIDIA's optimized models for production deployments with low latency and high throughput.
Tech: NVIDIA AI Models, LangChain, LangServe, Python, Streamlit
Object Tracking System
Real-time object tracking using OpenCV with Channel and Spatial Reliability Tracking (CSRT). Practical computer vision application for surveillance, sports analysis, or any scenario requiring robust object tracking.
Tech: OpenCV, CSRT, Python
Food Classification with VGG-16
Deep learning project using transfer learning with VGG-16 for automated food image classification. Demonstrates the power of pre-trained models and transfer learning for domain-specific tasks.
Tech: TensorFlow, Keras, VGG-16, NumPy, Matplotlib, Pandas, OpenCV
AI Lecture Transcriber
Convert YouTube videos into detailed study notes across various subjects including OpenCV, Machine Learning, LLMs, Data Science & Statistics, and Generative AI. A practical tool for students who prefer reading to watching videos.
Tech: Streamlit, LangChain, YouTube API
Multi-Language Invoice Generator
Leverage Google's Gemini Vision Model to extract and generate invoices in multiple languages. Perfect for businesses operating internationally or handling diverse linguistic requirements.
Tech: Google Gemini Vision, Streamlit, Python
Project Generator Tool
Tired of staring at blank screens? This tool generates personalized data project ideas based on your job title, favorite tools, and industry. It creates detailed project suggestions with timelines and skill requirements to help bring ideas to life.
Tech: Google Gemini, Streamlit, Pandas, Matplotlib, Plotly
Ollama UI
Interactive UI for running and managing models locally using Ollama. Demonstrates how to create user-friendly interfaces for local LLM deployment, giving you full control over your AI models.
Tech: Ollama, Streamlit, Python, OpenAI-compatible APIs
🏆 Recognition & Credentials
- 📄 Published Research: IEEE OTCON 2025 - "AI-Associate: A Lightweight Architecture for Conversational Agents"
- 🎓 Education: B.E. from Priyadarshini Engineering College, Higna Road, Nagpur (RTM Nagpur University)
- 🏅 Certifications:
- IBM AI Ladder Framework
- DeepLearning.AI - Intro to TensorFlow for AI
- 4+ total technical certifications
- 🌟 Open Source: 2+ significant contributions to community projects
- 💻 Portfolio: 15+ production-ready AI/ML projects across various domains
- 🏆 Community: 500+ connections on LinkedIn, active on Kaggle and GitHub
- 📝 Technical Writing: Regular contributor on Dev.to and Medium
💡 What I've Learned About AI Engineering
After building these projects, here's my hard-earned wisdom:
1. Start with the Problem, Not the Tech
It's tempting to think "I want to use GPT-4" or "I should try LangChain." Resist. Start with a real problem, then find the appropriate tools.
2. Deployment is Half the Battle
A Jupyter notebook is not a product. If users can't access it, it doesn't matter how good the model is. Learn Docker, learn APIs, learn DevOps.
3. Data Quality > Model Complexity
I've seen a simple model with clean, relevant data outperform a complex ensemble on messy data every single time.
4. Context Matters More Than You Think
Building for Indian languages taught me that cultural context, regional variations, and user expectations are as important as technical accuracy.
5. Document Everything
Future you will thank present you. Write READMEs, add comments, create architecture diagrams. Your portfolio is your documentation.
🎯 What's Next?
I'm currently exploring:
- Edge AI: Running models on resource-constrained devices
- Multimodal fusion: Better combining vision, language, and audio
- AI safety: Making models more reliable and interpretable
- Developer tools: Building better experiences for AI engineers
🤝 Let's Connect!
I'm eager to contribute to impactful projects that drive positive societal change. My focus lies at the intersection of data science and machine learning, and I'm a committed learner who thrives on engaging with a diverse community of data professionals, fostering a spirit of knowledge-sharing.
Building AI systems that matter requires collaboration and community. I'd love to:
- Discuss these projects in detail
- Collaborate on open-source initiatives
- Share knowledge about AI engineering
- Learn from your experiences
- Explore the exciting possibilities that await in the dynamic world of technology
Find me here:
- 🌐 Portfolio: aadyamadankar.life
- 💻 GitHub: @Aadya-Madankar
- 💼 LinkedIn: Aadya Madankar (500+ connections)
- 📝 Medium: @aadyamadankar1099
- 💬 Dev.to: @aadya_madankar_6dc52aeee1
- 📊 Kaggle: aadyamadankar
- 🚀 Devpost: Aadya1603
- 📧 Email: aadyamadankar1099@gmail.com
- 📍 Location: Nagpur, Maharashtra, India
🎬 Final Thoughts
Your portfolio isn't just a collection of projects—it's a demonstration of how you think, what you value, and what you're capable of building.
Mine shows that I care about accessibility, education, and practical impact. It demonstrates technical depth across the AI stack while staying grounded in real-world applications.
What does yours say about you?
If you're building your own AI engineering portfolio, remember:
- Pick projects that genuinely interest you
- Solve real problems, even if small ones
- Document your process, not just your results
- Share what you learn along the way
The best AI engineers aren't just prompt engineers or model fine-tuners—they're problem solvers who happen to use machine learning as a tool.
Now go build something amazing! 🚀
What projects are you working on? Drop a comment below—I'd love to hear what you're building!
Top comments (0)