• 🎓 B.Tech Artificial Intelligence · Anurag University, Hyderabad (CGPA: 8.07 · Graduating 2026)
• 🔍 Open to: Data Analyst · Data Scientist · AI Engineer · Python Developer · ML Developer
• 🛠️ I build end-to-end data products — raw data → EDA → ML models → deployed apps
• 📧 mloukikreddy@gmail.com · 📍 Hyderabad, India

Python · FastAPI · LangChain · FAISS · sentence-transformers · Llama 3.3 · Gemini 2.0 · Cohere · DuckDuckGo · Docker · HuggingFace Spaces

Upload any PDF, DOCX, PPTX, or TXT and chat with it using 3 AI models simultaneously. Built a complete RAG pipeline from scratch — ingestion, vector embeddings, semantic retrieval, parallel multi-LLM inference, and response fusion.

How it works:

User Uploads Document (PDF / DOCX / PPTX / TXT)
           ↓
   Text Extraction (pypdf, python-docx)
           ↓
   Split into 500-token chunks (50-token overlap)
           ↓
   Embed via all-MiniLM-L6-v2 (~90MB, runs locally)
           ↓
   Store in FAISS vector database (per-session UUID)
           ↓
   User asks a question
           ↓
   Semantic search → Top 4 relevant chunks
         +
   DuckDuckGo web search → Top 3 results
           ↓
   Combined context → 3 LLMs in parallel (ThreadPoolExecutor)
   [Llama 3.3]  [Gemini 2.0 Flash]  [Cohere Command-R]
           ↓
   Groq/Llama merges all 3 → Final answer returned

Key design decisions:

• Session isolation: UUID per user → separate FAISS index → no cross-user data leakage
• Chat history: Last 5 turns injected into every prompt → follow-up questions work correctly
• Lazy loading: Models load on first request, not startup → fits within Render's 512MB RAM limit
• $0 cost: 100% free APIs + local embeddings — no OpenAI, no Pinecone

• 🔗 GitHub Repository
• 🌐 Live Demo — HuggingFace Spaces

Python · DenseNet121 · Fusion Layer · LightGBM · TensorFlow/Keras · Jupyter

Advanced multi-modal deep learning system for DR severity classification. Dual DenseNet streams process two image modalities independently, outputs fused via a custom fusion layer, then passed to a classical ML classifier for final prediction.

How it works:

Fundus Image  +  OCT Retinal Image
      ↓                  ↓
 DenseNet Stream A   DenseNet Stream B
 (spatial features)  (texture features)
      ↓                  ↓
      └──── Fusion Layer ────┘
                  ↓
       Combined feature vector
                  ↓
       ML Classifier (LightGBM)
                  ↓
   5-class DR severity prediction:
   [Normal] [Mild] [Moderate] [Severe] [Proliferative DR]
                  ↓
   Evaluation: Accuracy · Precision · Recall · F1 · Confusion Matrix

Key design decisions:

• Dual DenseNet streams: Each modality (Fundus vs OCT) carries different clinical signals — parallel streams capture both independently before fusion
• Fusion Layer: Concatenates feature vectors from both streams → richer representation than single-stream
• DenseNet over VGG: Dense connections reuse features across layers → better gradient flow, more efficient on small medical datasets
• LightGBM classifier head: Faster inference + interpretable feature importance vs. full softmax DNN head

• 🔗 GitHub Repository
• 

Python · SQL · SQLite · Pandas · Excel · Power BI · Olist Dataset

End-to-end analytics pipeline on real Brazilian e-commerce transactions (Olist dataset) — from raw CSV ingestion to a 3-page interactive Power BI dashboard with actionable business KPIs.

How it works:

Raw Olist Dataset (Kaggle · multi-table relational data)
           ↓
   Excel → initial preprocessing & data validation
           ↓
   Python (Pandas) → data cleaning, merging tables,
   feature engineering (delivery time, review mapping)
           ↓
   SQLite (ecommerce.db) → structured storage
   SQL queries → KPI aggregations:
   [Total Revenue] [Total Orders] [AOV]
   [Avg Delivery Time] [Review Scores]
           ↓
   EDA Notebook (ecommerce_eda.ipynb)
   → Monthly revenue trends
   → Payment method distribution
   → Delivery performance vs. review score correlation
           ↓
   Power BI → 3-page interactive dashboard
   [Page 1: Executive Overview]
   [Page 2: Detailed Insights]
   [Page 3: Advanced Analytics]

Key design decisions:

• SQLite as warehouse: Lightweight, file-based DB — no server setup, reproducible, commitable to repo as ecommerce.db
• Multi-page Power BI: Separate pages for exec summary, product drill-down, and delivery analytics — prevents overcrowding a single canvas
• Delivery-satisfaction correlation: Surfaced that delivery delays directly tank review scores — key actionable insight for business

• 🔗 GitHub Repository

Python · VGG16 · LightGBM · TensorFlow/Keras · OpenCV · Google Colab · Kaggle

Hybrid ML pipeline for automated DR screening from retinal fundus images (~5,000 images). Pretrained CNN for feature extraction + gradient boosting classifier — high accuracy at low inference cost.

How it works:

Retinal Fundus Images (~5,000 images · Kaggle dataset)
           ↓
   Preprocessing: resize → normalize → augmentation
           ↓
   VGG16 (pretrained on ImageNet)
   → Remove top classification layers
   → Extract deep feature vectors from conv layers
           ↓
   Feature vectors → LightGBM classifier
           ↓
   5-class DR severity prediction
           ↓
   Accuracy: 99%+ · Sample confidence: 99.50%
   Evaluation: Precision · Recall · F1 · Confusion Matrix

Key design decisions:

• Hybrid DL + ML: VGG16 extracts rich spatial features; LightGBM classifies faster + more interpretably than a full DNN softmax head — better for small medical datasets
• Transfer learning: ImageNet pretrained weights give strong edge detection and texture features critical for retinal vessel analysis
• Kaggle API ingestion: Dataset too large to commit → Kaggle API setup included for fully reproducible Colab runs

• 🔗 GitHub Repository
• 

Node.js · Express.js · HTML · CSS · JavaScript · Pug · Weather API · Render

Full-stack real-time weather application. Search any city → get live temperature, humidity, wind speed, and conditions instantly. Deployed and live on Render.

How it works:

User enters city name in browser UI
           ↓
   Express.js server receives GET request
           ↓
   Server-side call to Weather API (live data)
           ↓
   API returns JSON:
   [temp] [humidity] [wind speed] [condition]
           ↓
   Express parses → renders via Pug template
           ↓
   Dynamic HTML returned → browser displays live data

Key design decisions:

• Server-side API call: API key stays on Node/Express server, never exposed to client browser — prevents key theft
• Pug server-side rendering: Fast initial load, no client-side JS framework overhead for a simple data-display app
• Render deployment: Free tier persistent process → app stays live at stable URL

• 🔗 GitHub Repository
• 🌐 Live Demo

MEAN Stack · MongoDB · Express.js · Angular · Node.js · Render

Full-stack location discovery and review platform on the MEAN stack. Users explore places, sign in securely, and add reviews. Live on Render.

How it works:

User visits Home page
           ↓
   Angular frontend → HTTP call to Express REST API
           ↓
   Express routes → controllers → MongoDB queries
           ↓
   MongoDB returns location/review documents
           ↓
   Angular renders dynamic UI (no full page reload)
           ↓
   Auth flow: Sign In → session token → write access
           ↓
   Authenticated user submits review
   → POST /api/reviews → persisted in MongoDB

Key design decisions:

• MEAN stack: Single language (JavaScript) across frontend, backend, and DB layer → faster dev, consistent data models
• REST API-first: Separate app_api/ layer decouples backend from frontend → reusable, testable
• MongoDB: Schema-flexible NoSQL fits location/review data with variable fields — no rigid table migrations

• 🔗 GitHub Repository
• 🌐 Live Demo

Data & Analytics

Machine Learning & AI

Web & Deployment

• 📧 Email: mloukikreddy@gmail.com
• 💼 LinkedIn: mekala-loukik-reddy-a717b4287
• 🐙 GitHub: mloukikreddy

"Build. Analyze. Deploy."