Portfolio - Sarthak Sharma

Sarthak Sharma

DS & GENAI DEV

Resume

1

SQL Project

1

Python Project

3

Machine Learning Projects

3

Deep Learning Projects

3

GenAI Projects

About Me

I'm Sarthak Sharma, a Computer Science (AI) graduate from Bennett University, passionate about building AI systems that solve real problems. My focus areas are Generative AI, Data Science, and backend development, with hands-on experience across LLM fine-tuning, RAG pipelines, recommendation systems, and machine learning workflows.
Through internships at Spacekayak, Star Vacays, and Eleva AI, I've worked with real-world data, deployed models on AWS, and built production APIs using Python and FastAPI. I've also built projects like NyayaSaathi, an offline legal-aid bot using RAG, and agentic AI tools for data analysis.
My strength is shipping fast, taking an idea from concept to a working, deployed solution. I'm always learning, whether it's a new framework, a research paper, or a tough debugging problem. I'm looking to grow as an AI/ML engineer and contribute to teams building practical, impactful AI products.

Key Skills

Python

C++

SQL

Machine Learning

Deep Learning

Data Analysis

EDA

Pandas

NumPy

scikit-learn

FastAPI

REST APIs

AWS (EC2

GPU Instances)

Git

Adaptability

Communication

Critical Thinking

My Projects

WatchTower — NFT Scam Detection System

Domain/Function: Deep Learning / Blockchain Security

E-commerce Chatbot

Domain/Function: E-commerce / Retail

Car Damage Detection

Domain/Function: Computer Vision — Insurance / Automotive

RAG Research Assistant with Source-Verified Citations

Domain/Function: Financial & Real Estate News Research

NyayaSaathi — Offline Legal Aid Assistant

Domain/Function: Legal Tech / GovTech

NSE/BSE Stock Advisor

Domain/Function: FinTech / Stock Market Analysis

Expense Tracking System

Domain/Function: Personal Finance

AtliQ Hotels — Hospitality Data Analysis

Domain/Function: Hospitality / Business Intelligence

Health Insurance Premium Predictor

Domain/Function: InsurTech

Caltech-101 Image Classifier

Domain/Function: Computer Vision

Credit Risk Modelling — CredIQ Finance

Domain/Function: Machine Learning / FinTech

My Experience

Superteam India — Web3 & AI Contributor
2024 – Present

Active contributor in the Superteam India ecosystem, participating in multiple Web3 and AI bounties across the Solana developer community.
Secured 3rd place all over India in a Solana hackathon (Watchtower), winning a cash prize for building an AI-powered Web3 solution.
Engage regularly with the Superteam DAO community, collaborating on projects and helping peers debug technical problems.

Star Vacays, Remote — Data Science & AI Developer Intern
Jul 2024 – Mar 2025

Built an AI-powered recommendation system for a vacation booking platform, personalizing destination suggestions to improve user engagement.
Worked with real-world production data, handling cleaning, preprocessing, and feature engineering to prepare messy datasets for modeling.
Applied data science techniques including EDA and model building to support the platform's personalization features.

Spacekayak, Remote — Machine Learning Engineer Intern
Oct 2023 – Jan 2024

Built an AI headshot generation system using generative AI models to produce professional headshots from user inputs.
Used Stability AI's Stable Diffusion and Stable Video Diffusion to create a pipeline for Flipkart product content.
Fine-tuned the Zephyr LLM for Flipkart product comparison and summarization, building custom LLM workflows before frameworks like LangChain were widely adopted.
Ran generative AI and LLM workloads on high-compute GPU-backed AWS instances.

Awards & Certificate

Python: Beginner to Advanced For Data Professionals

SQL for Data Science

Math and Statistics For AI, Data Science

Master Machine Learning for Data Science & AI: Beginner to Advanced

Deep Learning: Beginner to Advanced

Natural Language Processing

Gen AI to Agentic AI with Business Projects

Generative AI for Everyone

AI and Climate Change

Let's Connect

Feel free to get in touch with me. I am always open to discussing new projects, creative ideas or opportunities to be part of your visions.

Download Resume

Resume

Mail Me

[email protected]

https://www.linkedin.com/in/sarthaksharma1411/

Call Me

+91 9662464745

GitHub

https://github.com/sarthak-here

Deep Learning Project

WatchTower — NFT Scam Detection System

Domain/Function: Deep Learning / Blockchain Security

AI-powered NFT scam detection system for Solana blockchain. Combines BERT and Random Forest to scan wallets, identify fraudulent NFTs, and recommend burn actions with 99.8% accuracy and real-time confidence scores.

View Project
Engagement

View Project
on GitHUB

Other Tools

BERT

RandomForest

Next.js

AWS Lambda

AWS EC2

API Gateway

Shyft API

WatchTower frontend — Solana wallet connection interface showing NFT scan dashboard with real-time scam detection, spam flagging, and burn recommendations for suspicious assets.

Real-world impact — VCs and founders who fell victim to Solana NFT scams, highlighting the problem WatchTower was built to solve.

Project Details

WatchTower is an AI-powered security system that protects Solana blockchain users from NFT scams and spam. Users connect their Solana wallet and the system scans all holdings, analyzing each NFT against a custom-trained model combining BERT for natural language understanding and Random Forest for classification. The model was trained on real Solana NFT data and achieves 99.8% accuracy in distinguishing scam, spam, and legitimate NFTs. Flagged assets are surfaced with confidence scores and burn recommendations, giving users clear action steps. The backend is fully serverless on AWS using API Gateway, Lambda, and EC2, with NFT metadata fetched via the Shyft API. The frontend is built in Next.js for a fast, responsive user experience. WatchTower acts as a foundational wallet security layer — real-time, automated, and requiring no manual review from the user.

Project video

GenAI Project

E-commerce Chatbot

Domain/Function: E-commerce / Retail

A multi-agent e-commerce assistant that handles FAQ RAG, live product search, review-based Q&A, comparisons, recommendations, and image-based visual search across 8,765 products and 20,505 customer reviews.

View Project
Engagement

View Project
on GitHUB

Other Tools

Streamlit

Groq

ChromaDB

SQLite

semantic-router

sentence-transformers

CLIP

System design showing semantic routing between FAQ RAG, live SQL search, review analysis, recommendations, comparisons, and visual search.

Upload a product image to retrieve visually similar catalogue items using CLIP image embeddings.

Project Details

Built a multi-agent e-commerce chatbot that routes each shopper query to the appropriate capability: FAQ retrieval, live SQLite product search, customer-review Q&A, product comparison, recommendations, and CLIP-based visual search. The application uses semantic-router, ChromaDB, Groq, Streamlit, SQLite, and sentence-transformers. It supports a catalogue of 8,765 products across 10 categories, 20,505 reviews, and 8,765 indexed product images.

Deep Learning Project

Car Damage Detection

Domain/Function: Computer Vision — Insurance / Automotive

An end-to-end deep learning system that classifies car damage from photos into 6 categories (front/rear breakage, crushed, normal). Built with PyTorch, tuned via Optuna, and served through both a Streamlit UI and a FastAPI endpoint, reaching 85.2% validation accuracy.

View Project
Engagement

View Project
on GitHUB

Other Tools

PyTorch

Streamlit

FastAPI

Optuna

ResNet50

The Streamlit UI: upload a photo of a car and get an instant prediction across the six damage categories, with confidence scores.

model_helper.py — the inference module: defines the fine-tuned ResNet50 classifier (layer3/layer4 unfrozen, custom 256-unit head), applies the image preprocessing pipeline, and runs a forward pass to predict the damage class.

Project Details

Car Damage Detection is a vehicle damage classifier for insurance claims and inspection workflows. Given a photo of a car, it predicts one of six damage classes: front/rear breakage, front/rear crushed, or front/rear normal.

- I compared six model variants on a ~2,300-image dataset: a plain CNN, a CNN with batch norm and dropout, transfer learning on EfficientNet-B0, and progressively fine-tuned ResNet50 variants. The final model came from a 100-trial Optuna search across five architectures (ResNet50, ConvNeXt-Tiny, EfficientNet-B0, DenseNet121, MobileNetV3-Large), jointly tuning learning rate, dropout, optimizer, and weight decay.

- The winning model, a fine-tuned ResNet50 with a custom classifier head, reached 85.2% weighted F1 on held-out validation. The weakest class, Rear Breakage, is most often confused with Rear Crushed.

- Served via a Streamlit UI and a FastAPI endpoint.

GenAI Project

RAG Research Assistant with Source-Verified Citations

Domain/Function: Financial & Real Estate News Research

A retrieval-augmented generation assistant that answers questions about news articles and cites the exact passages behind every answer. Built so citations come from retrieval metadata rather than model output, and tested adversarially to confirm it refuses rather than fabricates.

View Project
Engagement

View Project
on GitHUB

Other Tools

LangChain

ChromaDB

HuggingFace Sentence Transformers

Groq

Streamlit

The Streamlit interface answering a question against two indexed CNBC articles. The response reports only the figures actually present in the source text rather than asserting one it cannot support, and lists both source URLs beneath the answer.

Asked for Bitcoin's price against a corpus of mortgage-rate articles, the assistant refuses and explains what the indexed documents actually cover, instead of generating a plausible figure.

Project Details

The system scrapes article URLs, chunks and embeds the content into a persistent Chroma vector store, retrieves the relevant passages for each query, and generates answers using GPT-OSS 120B on Groq behind a Streamlit interface.

The focus was citation reliability rather than answer fluency.

Citations are derived from the metadata of retrieved chunks rather than parsed from the model's output, keeping the citation path independent of the component it verifies. Retrieved context is assembled against an explicit budget rather than passing top-k straight into the prompt, so listed sources reflect only what the model actually received. Grounding is validated adversarially by querying with questions the corpus cannot answer, confirming it declines rather than invents.

Implemented on LangChain 1.x with LCEL composition, with ingestion, retrieval and serving as independent modules over one set of lazily loaded components.

GenAI Project

NyayaSaathi — Offline Legal Aid Assistant

Domain/Function: Legal Tech / GovTech

NyayaSaathi is a 100% offline legal aid assistant powered by Gemma 4. It identifies applicable Indian laws from 500+ provisions, explains your rights in Hindi/English/Hinglish, drafts formal complaints, and guides ITR filing, all without your data leaving your device.

View Project
on GitHUB

Other Tools

Gemma 4

Ollama

LangChain

ChromaDB

FastAPI

Whisper

fpdf2

The NyayaSaathi landing page: describe your legal problem in Hindi, English, or Hinglish and get instant analysis. Highlights 500+ laws indexed, 64 legal documents, 100% offline operation, and quick-access categories like Police Complaint and Cyber Crime.

The core analysis flow: describe your situation on the left (with voice input and demo scenarios in Hindi/English/Hinglish), and Gemma 4 generates a legal analysis on the right, entirely offline, showing applicable laws and your rights in plain language.

Project Details

NyayaSaathi is a free, 100% offline legal aid assistant for underserved Indians who can't afford legal counsel. Powered by Gemma 4 running locally via Ollama, it never sends your data to any server.

- Describe your situation in Hindi, English, or Hinglish and it identifies applicable Indian laws from a RAG-indexed knowledge base of 500+ laws and provisions spanning criminal law, labour rights, consumer protection, family law, property, and cyber crime. It explains your rights in plain language, answers contextual follow-up questions, and can AI-draft formal complaints across 10 types (FIR, consumer, cyber crime, RTI, POSH, and more), exportable to PDF.

- It also includes an ITR Filing Advisor: upload a bank statement and get the correct ITR form, an Old vs New tax regime comparison, and a personalized deduction checklist (80C, 80D, HRA, NPS, and more).

- Built with FastAPI, ChromaDB, LangChain, and offline Whisper for voice input. Built for the Kaggle Gemma 4 Hackathon 2025.

Machine Learning Project

NSE/BSE Stock Advisor

Domain/Function: FinTech / Stock Market Analysis

An AI-powered Indian stock market advisor that screens ~164 NSE/BSE stocks daily for the top 30 movers, generates BUY/SELL/HOLD signals from a weighted technical engine plus Gemma 4 news sentiment, and presents it all in an interactive Streamlit dashboard.

View Project
on GitHUB

Other Tools

yfinance

pandas-ta

Streamlit

Plotly

Gemma 4

Claude

APScheduler

Deep Dive position sizing panel: enter capital and risk % to get exact share count and position value, alongside an AI-generated News Impact summary (bullish/neutral/bearish breakdown) and one-click Add to Watchlist with pre-filled alert levels.

Deep Dive candlestick chart for a screened stock, with EMA20/50, RSI, Bollinger Bands, and volume overlays. The side panel shows the BUY/SELL/HOLD signal, confidence %, live price, and full trade levels (entry, stop loss, targets 1 & 2).

Project Details

NSE/BSE Stock Advisor is an AI-powered dashboard for Indian equities. A daily screener scans ~164 NSE/BSE stocks and ranks the top 30 by volume spike, ATR%, and momentum, with a live monitor swapping in stronger movers every 15 minutes.

- Signals come from a weighted multi-indicator engine (EMA21/50/200 trend alignment, MACD crossovers, RSI, ADX trend strength, context-aware Bollinger Bands), combined with AI news sentiment from Gemma 4 (or Claude), which fetches and analyzes headlines for all 30 stocks in a single call. Agreement between technicals and news boosts confidence; contradiction penalizes it.

- Features include a sector heat map, per-stock deep dives with candlestick charts and position sizing, a searchable universe of all 164 stocks, a watchlist with price alerts and live P&L, and a walk-forward backtester reporting Win Rate, Profit Factor, Sharpe Ratio, Max Drawdown, CAGR, and a full trade log.

- Built with Streamlit, yfinance, pandas-ta, and Plotly.

SQL Project

Expense Tracking System

Domain/Function: Personal Finance

Full-stack expense tracker built with FastAPI and MySQL: log daily expenses by amount, category, and notes, view any date's spending, and see category-wise breakdowns with percentages and charts over a custom date range, via a Streamlit frontend.

View Project
on GitHUB

Other Tools

FastAPI

MySQL

Streamlit

Pandas

pytest

The Add Expense tab: an editable table of logged expenses pulled live from MySQL, each row showing amount, category (via dropdown), and notes, with inline add/update controls.

The Analytics tab: category-wise spending breakdown with totals and percentages in a table, plus a bar chart visualizing the same data across a selected date range.

Project Details

Expense Tracking System is a full-stack Python app for logging and analyzing personal expenses. A FastAPI backend exposes endpoints to add, fetch, and analyze expenses, backed by a MySQL database (mysql-connector-python) with a simple expenses table (date, amount, category, notes).

- The Streamlit frontend has two tabs: one to add/update expenses for a given date, and an analytics tab that queries a date range and renders a category-wise spending breakdown with percentages and a bar chart, powered by Pandas.

- The project also includes a custom Python logging setup and a pytest suite covering the database helper functions, so the data layer is unit-tested independently of the API.

- Tech stack: FastAPI, Uvicorn, Pydantic, MySQL, Streamlit, Pandas, pytest.

Python Project

AtliQ Hotels — Hospitality Data Analysis

Domain/Function: Hospitality / Business Intelligence

A data analysis of AtliQ Hotels, a fictional 25-property Indian hotel chain, digging through 134,000+ booking records across 4 cities to find what drives occupancy and revenue, and where the business is leaving money on the table.

View Project
on GitHUB

Other Tools

Pandas

NumPy

Matplotlib

Jupyter Notebook

Occupancy rate by city across AtliQ's four markets: Delhi leads at 61.5%, ahead of Hyderabad and Mumbai (~58%) and Bangalore (56.3%), despite not being the top revenue earner.

Total revenue by property: Atliq Exotica leads at roughly Rs. 22 crore while Atliq Seasons trails nearly 5x behind at the bottom, flagging it for further investigation.

Project Details

AtliQ Hotels Analysis works through booking and occupancy data for a fictional 25-property Indian hotel chain across Delhi, Mumbai, Hyderabad, and Bangalore (May-Aug 2022, 134,590+ bookings), using Python and Pandas.

- The raw data was messy: negative guest counts, revenue outliers, capacity mismatches, and 57.8% missing ratings. Each issue was cleaned with a documented, justified method (e.g. mean + 3*std dev threshold for revenue outliers) rather than blanket imputation.

- Key findings: weekend occupancy (72.34%) far outpaces weekday (50.88%), pointing to underused weekday business travel. Delhi has the highest occupancy (61.51%) but the lowest revenue among the four cities, revealing a pricing gap. One property, Atliq Seasons, earns 5x less than the top performer, Atliq Exotica. Direct bookings account for 40.9% of volume, the largest single channel, ahead of any OTA.

- Tools: Python, Pandas, NumPy, Matplotlib, Jupyter Notebook.

Machine Learning Project

Health Insurance Premium Predictor

Domain/Function: InsurTech

A machine learning web app that predicts annual health insurance premiums from 13 demographic, lifestyle, and medical features. Trained on 50,000 records with age-based model segmentation, it reaches R² > 0.98 on both customer segments, deployed via Streamlit.

View Project
on GitHUB

Other Tools

scikit-learn

XGBoost

Pandas

statsmodels

Streamlit

The 12-field input form for the Health Insurance Premium Predictor: age, dependants, income, genetical risk, insurance plan, employment, gender, marital status, BMI, smoking status, region, and medical history.

Live prediction for a 42-year-old, Gold plan, diabetic profile: the app routes to the age-appropriate model (XGBoost for age>25) and returns an estimated annual premium of Rs. 25,024.

Project Details

Health Insurance Premium Predictor estimates annual premiums from 13 features (age, BMI, smoking status, medical history, income, plan tier) on a 50,000-record dataset.

- A single XGBoost model reached R²=0.978 overall, but ~30% of predictions for customers aged 25 and under had >10% error - young customers' premiums are driven by genetic risk and plan tier rather than medical history, a pattern the unified model missed.

- Fix: two specialized models split by age. Linear Regression handles age <=25 (R²=0.989, 2.14% extreme errors); a tuned XGBoost (RandomizedSearchCV, 3-fold CV) handles the rest (R²=0.995, 0.32% extreme errors) - cutting extreme errors by ~93-99% vs. baseline.

- Pipeline: outlier removal, medical-risk scoring from parsed diagnosis strings, VIF-based multicollinearity checks, segment-specific MinMax scaling. Deployed as a Streamlit app with a 12-field input form.

- Tech: Python, scikit-learn, XGBoost, Pandas, statsmodels, Streamlit.

Deep Learning Project

Caltech-101 Image Classifier

Domain/Function: Computer Vision

A PyTorch image classifier for all 101 Caltech-101 object categories, using transfer learning across six ImageNet-pretrained backbones tuned with Optuna. The final ResNet50 model reaches 93.84% test accuracy, served through a Streamlit app.

View Project
on GitHUB

Other Tools

PyTorch

torchvision

Optuna

Streamlit

The model correctly classifies the uploaded photo as "elephant" at 96.09% confidence, with a top-5 class probability bar chart comparing it against the next most likely categories.

A second test image, a kangaroo, is also classified correctly at 91.72% confidence, showing the model generalizes across different object categories rather than overfitting to one class.

Project Details

Caltech-101 Classifier is a PyTorch transfer-learning image classifier covering all 101 Caltech-101 categories (~9,146 images, natively imbalanced 32-800 per class).

- Six ImageNet-pretrained backbones (ResNet18/34/50/101, EfficientNet-B0, InceptionV3) were compared across three rounds of Optuna search (TPE sampler, median pruner), tuning backbone, freeze-vs-fine-tune, dropout, optimizer, LR, weight decay, batch size. ResNet50 won every round.
- Final run adds cosine LR schedule, gradient clipping, mixed-precision training, and label smoothing, fine-tuned 25 epochs end-to-end. Reaches 93.84% accuracy on a genuinely held-out 20% test split (macro-F1 0.917), reached after catching and correcting an earlier optimistic estimate that mixed validation-split scoring with cherry-picked best-of-N runs.
- Ships with a Streamlit app to upload a photo and see the prediction plus a top-5 probability chart, a pytest suite, and CI on every push.
- Tech: PyTorch, torchvision, Optuna, Streamlit

Machine Learning Project

Credit Risk Modelling — CredIQ Finance

Domain/Function: Machine Learning / FinTech

End-to-end loan default prediction system trained on 50k records from 3 data sources. Logistic Regression with SMOTETomek and Optuna tuning achieves AUC=0.98, Gini=0.96. Outputs credit score (300–900) and risk rating.

View Project
on GitHUB

CredIQ Finance credit risk dashboard — input form for borrower details including age, income, loan amount, tenure, DPD, and delinquency ratio to predict default probability and credit score.

main.py — Streamlit app source showing input form layout, loan-to-income ratio auto-calculation, and predict function call returning default probability, credit score, and rating.

Project Details

Built a credit risk prediction pipeline merging customer demographics, loan details, and bureau data (50k records). Performed stratified train-test split before EDA to prevent data leakage. Cleaned missing values, removed processing fee outliers using business rules, and fixed data entry errors. Engineered three key features: Loan-to-Income ratio, Delinquency Ratio, and Average DPD per Delinquency. Used VIF analysis to remove multicollinear features and WOE/IV scoring to retain only high-signal predictors. Addressed 8.6% class imbalance with SMOTETomek. Compared Logistic Regression, Random Forest, and XGBoost across 4 modeling attempts. Tuned final Logistic Regression with Optuna achieving AUC=0.98 and Gini=0.96. Evaluated using ROC curve, KS statistic, and decile rank ordering. Deployed a Streamlit app that computes default probability, maps it to a 300–900 credit score, and assigns a rating.

Already have an account? Sign in