HARSH RAJ

01.

About

I'm Harsh Raj, an MCA candidate focused on generative AI and AI backend systems. With hands-on experience in machine learning and API development, I thrive on creating impactful solutions that drive efficiency and insights in complex data environments.

02.

Experience

Feb 2025 - March 2025

AI/ML Intern – Automation Division · Tata Steel Limited

Jamshedpur (Onsite)

Built a time-series forecasting model for 8-hour PSD prediction in an industrial setting, using plant data to support hydrocyclone parameter optimization.
Processed 45 shift-wise plant records, addressing missing values and outliers, while benchmarking VAR, ARIMA, and LSTM models in Python and Keras to identify the optimal forecasting method.
Delivered the final LSTM model with an average Mean Absolute Error (MAE) of 1.38, enhancing accuracy by approximately 15% over VAR and 56% over ARIMA on the same dataset.

03.

Selected work

X

X Intelligence MCPCheck out

Developed an MCP (Model Context Protocol) server that connects LLM clients (Claude, ChatGPT) to the official X API v2 for reading, searching, and analyzing posts and profiles.
Implemented OAuth/bearer-token authentication, Zod-validated tool schemas, and optional AI-drafting integrations (OpenAI, Groq, Gemini, Anthropic), ensuring functionality in a production NitroCloud deployment.
Designed MCP tools enabling AI clients to retrieve X data through structured requests, streamlining API interactions.

GitHub

B

BrainOSCheck out

Constructed a local transformer-interpretability tool that runs Qwen2.5 models via PyTorch/Hugging Face, streaming real tokenizer, attention, hidden-state, and MLP-activation data to a React/WebGL 3D visualization over WebSocket.
Implemented a FastAPI backend with a custom autoregressive generation loop and KV-cache, supporting pause/resume/replay of inference steps across Apple MPS, CUDA, and CPU backends.
Developed a pytest suite to ensure robust testing and reliability of the tool.

GitHub

H

Hydro Cyclone PSD Simulator for Iron OreCheck out

Created a time-series forecasting pipeline to predict iron-ore feed particle-size distribution (PSD) for hydrocyclone operations, comparing VAR, ARIMA, and LSTM models on shift-wise plant data.
Preprocessed 45 shift-wise records through outlier removal and missing-value imputation, conducting an ADF stationarity check across all target columns before model fitting.
Achieved the lowest average Mean Absolute Error (1.38) with the LSTM model, improving predictions by 15% over VAR and 56% over ARIMA for 8-hour-ahead PSD forecasts.

Live Demo

04.

Skills & tools

Languages - PythonSQLHTMLJavaAI/ML - Machine LearningLLMsRAGAgentic AITransformer ModelsPyTorchScikit-learnKerasTime-Series ForecastingPrompt EngineeringTensorFlowLangChainDatabases - MySQLSQLiteChromaDBCloud & DevOps - DockerOther Technical Skills - NitroCloudMCPZodWebSocketREST APIs

05.

Education

SRM Institute of Science and Technology

Master Of Computer Applications (MCA)

2026 - 2028

Sarala Birla University

Bachelor of Computer Applications (BCA)

2023 - 2026

06.

Certifications

Professional Certificate Course in Generative AI and Machine Learning

E&ICT Academy, IIT Kanpur (MeitY & IIT Kanpur initiative) · 2026

Vocational Training Certificate - AI/ML Applications for Industries

Automation Dept., Tata Steel Ltd. (Prashikshan) · 2025

Internship Certificate - Machine Learning

Central Coalfields Ltd · 2025

Certificate of Achievement, Agentic AI Hackathon

SRMIST NitroStack.ai · 2026

07.

Open source

08.

Get in touch

Have a role or a project in mind? I'd love to hear from you.

sinhaharsh446@gmail.com +91 9142779224