Samarth Sugandhi
AI/ML Engineer building production-grade ML systems, RAG pipelines, and LLM-powered applications with FastAPI, PyTorch, and vector databases.
INDORE, IN
AI/ML Engineer building production-grade ML systems, RAG pipelines, and LLM-powered applications with FastAPI, PyTorch, and vector databases.
Navigating mutual fund documents is tedious — investors and advisors spend hours digging through PDFs for specific information. This RAG assistant reads mutual fund documents, converts them into searchable vector embeddings, and returns accurate answers using semantic search.
The system implements a production-ready retrieval pipeline:
Ingests 100+ mutual fund PDFs with smart chunking strategies for accurate retrieval without context loss.
Answer quality was rigorously measured using the RAGAS evaluation framework:
These scores demonstrate strong retrieval accuracy and response grounding against the source documents.
The application includes a FastAPI backend connected to ChromaDB vector store and a clean Streamlit frontend for interactive document querying.
Real-time object detection is critical for surveillance, autonomous systems, and industrial inspection. This system provides end-to-end detection with live video stream support and cloud deployment.
YOLOv8 was trained on 5,000+ annotated images achieving 92% mAP@0.5 with optimized inference achieving <100ms latency per frame on live video streams via OpenCV.
The system was containerized with Docker and deployed on AWS EC2, with a Flask REST API supporting both image upload and real-time webcam detection modes.
Accurate credit risk assessment is crucial for financial institutions. This ML pipeline predicts loan default probability using engineered features and optimized gradient boosting, outperforming traditional approaches.
The model was trained on 50K+ records with extensive feature engineering:
Outperformed a logistic regression baseline by 12% on AUC-ROC after automated hyperparameter optimization with Optuna.
Built a Flask REST API for real-time risk scoring with input validation and request logging. Predictions are stored in PostgreSQL for downstream analytics and reporting.
Financial literacy is hard to access. FinChat runs a local Mistral LLM to answer finance-related questions with full conversation history — no cloud API costs, complete data privacy.
The chatbot uses Ollama to serve a Mistral model locally, delivering fast inference without external API dependencies. The entire conversation pipeline runs on-premise.
Database schema designed with SQLAlchemy ORM to support multiple users and store chat sessions in PostgreSQL. Full conversation history is preserved and queryable.
Professional experience & certifications.
ML Engineer Intern — Fine-tuned Mistral-7B with QLoRA, built data pipelines, mentored 8 interns.
AI/ML Intern — Built text classification models and recommendation engine prototypes.
Oracle Generative AI Professional Certificate — certified in GenAI architectures.
AWS Cloud Practitioner Essentials — cloud infrastructure and services fundamentals.
Content reached professionals at Citadel, Goldman Sachs, Millennium, Morgan Stanley, and JPMorgan Chase.
ICC Women's World Cup 2025-26 & LitChowk Festival (30,000+ attendees). Social Media Head — grew AIML page reach 40%.
Fine-tuned Mistral-7B using QLoRA on 15K+ domain-specific samples, improving response accuracy by 23% over the base model. Deployed production inference endpoint via FastAPI handling 500+ daily requests.
Architected an automated data preprocessing pipeline for PDF extraction, text chunking, and embedding generation — reduced manual data preparation effort by 60%.
Mentored a cohort of 8 interns on deep learning fundamentals (backpropagation, optimizers, loss functions) through structured weekly sessions, achieving 100% capstone completion rate.
Engineered text classification models using scikit-learn and TF-IDF to auto-categorize 10K+ content scripts, achieving 87% accuracy and reducing manual tagging effort by 40%.
Designed a recommendation engine prototype using collaborative filtering to surface content themes based on audience engagement data across 20+ media assets.
Oracle Generative AI Professional Certificate — validated expertise in generative AI architectures, prompt engineering, and LLM deployment strategies.
AWS Cloud Practitioner Essentials — covering cloud infrastructure, compute (EC2), storage (S3), networking, and security fundamentals.
ML × Finance content on LinkedIn reached professionals at Citadel, Goldman Sachs, Millennium, Morgan Stanley, and JPMorgan Chase — establishing thought leadership at the intersection of machine learning and quantitative finance.
Event Coordinator for ICC Women's World Cup 2025-26 and LitChowk Festival with 30,000+ attendees.
Social Media Head — Grew Acropolis AIML department page reach by 40% through data-driven content strategy.
Python, SQL, C++, JavaScript, Bash
PyTorch, TensorFlow, scikit-learn, Hugging Face, OpenCV
LLM Fine-tuning, LangChain, LoRA/QLoRA, RAG Pipelines, Prompt Engineering
FastAPI, Flask, Django, REST APIs, SQLAlchemy, Pydantic
PostgreSQL, MySQL, MongoDB, ChromaDB, FAISS, Pinecone, Redis
Docker, AWS (EC2, S3), Git, Linux, MLflow, W&B, CI/CD
OpenCV, YOLOv8, CNNs, Image Segmentation, Object Detection
NLTK, spaCy, Pandas, NumPy, Matplotlib, Streamlit
That's it, you've reached the end of my portfolio. Thanks for visiting :)
If you enjoyed the journey, let's make the sequel together.
B.Tech CSE (AI & ML) — Acropolis Institute, Indore | RGPV, Bhopal