Ewan John Dennis

Builder · Kochi, India

---

About

Jumping from one muddy puddle to another—getting my hands dirty with new tech, figuring out how things work under the hood, and trying to get models out of notebooks into actual working projects. CS undergrad at MEC Kochi who loves tinkering with agentic AI, predictive ML, and building cool stuff with code.

---

Experience

FlyRank AI — Machine Learning Intern Jul 2026 – Aug 2026

Built a content refresh prioritization model ranking decay risk across 175,205 pages from search analytics. Trained LightGBM and Random Forest classifiers achieving ROC-AUC 0.973 and PR-AUC 0.9997 under strict GroupKFold cross-validation. Delivered a calibrated 3-tier action playbook (IMMEDIATE_REFRESH, SCHEDULED_UPDATE, MONITOR) with reason codes. Python LightGBM scikit-learn DuckDB pandas Hugging Face

Infosys Springboard — AI Engineer Intern Feb 2026 – Apr 2026

Designed and deployed a multi-agent AI architecture with 3 parallel concurrent agents, reducing data-collection latency by up to 3x. Integrated FAISS vector search, Groq LLM synthesis, and PyTorch LSTM forecasting for end-to-end automated intelligence generation. Python FAISS Groq PyTorch Streamlit MongoDB Atlas

---

Selected Work & Projects

RTIIS — Real-Time Industry Insight & Strategic Intelligence System 2026

Autonomous multi-agent market intelligence platform using parallel workers, FAISS RAG, and PyTorch LSTM forecasting. Shipped with MongoDB Atlas persistence, PDF exports, and a Streamlit UI deployed on HuggingFace Spaces. Python Groq FAISS PyTorch MongoDB Atlas Streamlit

Kavach — AI Chargeback Decisioning System 2026

End-to-end chargeback decisioning system built for the Razorpay AI Buildathon. Combines XGBoost prediction $P(\text{win})$ with an EV-based deterministic policy layer, SHAP feature explainability, and hallucination-free representment generation bound to merchant telemetry. XGBoost SHAP FastAPI SQLite pandas

TrainCLI — End-to-End ML Training CLI 2025 – 2026

Published PyPI tool to train, evaluate, and export scikit-learn/XGBoost models directly from a single CLI command. Features an integrated MCP server component to expose training and inference to agentic workflows. Python scikit-learn XGBoost MCP PyPI

Huddle — Real-Time Campus Collaboration Platform 2026

Real-time platform for students to find and join active study sessions on campus with sub-100ms Firestore latency. 🥈 2nd Place at Techsprint (GDGC MEC) out of ~40 teams. React Firebase Tailwind CSS Vercel

---

Technical Capabilities

AI / ML         PyTorch · FAISS · Groq · Agentic Workflows · RAG · LSTM Forecasting · LightGBM · XGBoost · scikit-learn
Providers       Groq · Hugging Face · OpenRouter · Azure OpenAI · Anthropic SDK
Backend         Python · FastAPI · Node.js · Express.js · Spring Boot · PostgreSQL · MongoDB · DuckDB
Frontend        TypeScript · JavaScript (ES6+) · React.js · Next.js · Tailwind CSS · Firebase
DevOps & Cloud  Docker · Git / CI-CD · GCP · Vercel · HuggingFace Spaces · Streamlit Cloud