MediTune
Task: LLM Fine-Tuning
DomainMedical QA
ModelMistral-7B
End-to-end QLoRA fine-tuning pipeline adapting Mistral-7B-Instruct-v0.3 for closed-domain Medical Question Answering (MedQA).
Machine Learning · Data · AI
Nikhil Sharma
I build as a
AI/ML Engineer specializing in LLM fine-tuning, real-time ML systems, and production MLOps.
3 ML models deployed·
94.2% TypeScript·
Lighthouse SEO: 100·
Python·
Next.js·
TensorFlow·
3 ML models deployed·
94.2% TypeScript·
Lighthouse SEO: 100·
Python·
Next.js·
TensorFlow·
3 ML models deployed·
94.2% TypeScript·
Lighthouse SEO: 100·
Python·
Next.js·
TensorFlow·
3 ML models deployed·
94.2% TypeScript·
Lighthouse SEO: 100·
Python·
Next.js·
TensorFlow·
3 ML models deployed·
94.2% TypeScript·
Lighthouse SEO: 100·
Python·
Next.js·
TensorFlow·
3 ML models deployed·
94.2% TypeScript·
Lighthouse SEO: 100·
Python·
Next.js·
TensorFlow·
01 / Work
Featured projects
All on GitHubStreamSentinel
Task: Real-Time ML
LatencyLow
ModelsEnsemble
Real-time anomaly detection pipeline — Kafka → FastAPI → Isolation Forest + Autoencoder → WebSocket dashboard. Low end-to-end latency.
Fraud Detection MLOps Pipeline
Task: MLOps
Precision0.83
Latency12ms
Dataset284k
End-to-end credit card fraud detection pipeline with real-time FastAPI inference, Dockerized serving, automated CI/CD, and Evidently AI drift detection triggering automatic retraining. avg_precision ~0.83 on Kaggle transaction data.
Model Accuracy94%
AgriFuture India
Task: Full-Stack AI
Accuracy96.2%
Classes38
AI-powered agri-tech platform with crop recommendation, plant disease detection, market forecasting, and a digital twin module. Built as ML Training Lead on a 4-person capstone team; deployed on Render.
Model Accuracy96%
RAG-QA-System
Task: LLM/RAG
Retrieval Rate0.91
Embeddings768-dim
Local-LLM document Q&A system — no data leaves the machine. Multi-format ingestion, FAISS vector search, streaming responses, multi-model switching.
Model Accuracy91%
Deepfake Detector
Task: Computer Vision
F1 Score0.93
FPS30+
Real-time detection of AI-generated face manipulation with visual tampering explanations across images, video, and live webcam, built on EfficientNet-B4.
Model Accuracy93%
House Price Prediction Platform
Task: ML/Deployment
R²0.641
MAE1.2M
Property valuation platform using XGBoost and LightGBM with SHAP-based explainability, deployed as an interactive app on Hugging Face Spaces. XGBoost R² = 0.641 on the Bengaluru housing dataset.
Model Accuracy88%
02 / Experience
Experience
AI/ML Intern · Tejaskp AI Software, Vadodara
Develop and optimize machine learning models (classification, regression) for real-world use cases, applying hyperparameter tuning to improve prediction accuracy. Build end-to-end data preprocessing pipelines using Python, Pandas, and NumPy, including EDA, feature engineering, encoding, and normalization. Evaluate model performance using cross-validation, ROC-AUC, and precision-recall metrics.
Data Analytics Intern · Trainity (Remote)
Performed exploratory data analysis and statistical analysis on user data using Python and SQL to surface engagement insights. Built data preprocessing pipelines handling missing values, encoding, and normalization to prepare clean datasets for ML modeling.
03 / Skills
ML Architecture Stack
Data Layer
Pandas
NumPy
SQL
Kafka
Processing
Hugging Face
LangChain
QLoRA/PEFT
Model Layer
PyTorch
TensorFlow
XGBoost
LightGBM
Deployment
FastAPI
Docker
AWS ECS
GitHub Actions
04 / About
Education
B.Tech, Computer Science Engineering
Data Science specializationITM Vocational University, Vadodara2022–2026CGPA 9.24
Certifications
- Applied Artificial Intelligence (Microsoft & SAP)
- Python for Data Science & Machine Learning
- Data Science & Analytics (HP)
- Data Visualization with Business Intelligence (Tata Forage)
05 / Contact
Get in touch.
Currently open to AI/ML internship opportunities and entry-level roles. Reach out via any channel below.