India
I build AI systems that reason, act, and hold up in production — not just in notebooks. My path into AI started during my M.Sc. in Mathematics & Scientific Computing at MNNIT Allahabad: the same frameworks I was studying — linear algebra, probability, optimization — were powering the most transformative technology of our time. That crossover from pure math to real-world problem solving pulled me in, and it still shapes how I think about every system I build. Here's something I believe strongly: not every problem needs a large language model. Before reaching for GPT or Llama, I ask — can a task-specific ML model solve this more reliably and interpretably? When LLMs are the right tool, I obsess over grounding them: reducing hallucination through RAG, applying guardrails, experimenting with prompting strategies, fine-tuning for task-specific behavior, and monitoring outputs through observability in production. The goal isn't the flashiest tool — it's the right one. What I enjoy most is designing multi-agent architectures — systems where specialized agents collaborate, validate each other's outputs, and adapt in real time. There's a unique satisfaction in orchestrating controlled autonomy, where the whole system is smarter than any single agent. One trend I'm genuinely excited about is MCP (Model Context Protocol). We used to write mountains of boilerplate to connect LLMs to external tools — email, files, APIs, codebases. MCP abstracts that away beautifully. The heavy lifting happens server-side; as a developer, you just connect. It's the kind of shift that makes agentic AI accessible to a much wider community, and I've been hands-on with 15+ MCP integrations in my current role. Outside work, I'm usually building side projects to test ideas I can't stop thinking about, or reading papers to stay close to where the field is heading. The best engineering decisions come from understanding research, and the best research questions come from hitting real production constraints. I care about hard, unsolved problems — regardless of domain. If it demands rigorous thinking, creative system design, and AI that works at scale, I want to be in that room. Open to full-time AI/ML Engineer and Data Scientist roles across India or remote. Let's connect — I'd love to exchange ideas on agentic AI, production ML, or anything at the intersection of math and intelligent systems.
Engineered an NL2SQL pipeline integrating LLMs with a semantic layer, supporting PostgreSQL, MySQL, MSSQL, Redshift, and BigQuery — achieved ~70% reduction in schema generation latency through batched parallel processing and ~60% lower inference cost via lightweight models with token-level cost tracking. Built a standalone RAG knowledge base featuring a 5-stage indexing workflow with support for 5 chunking strategies and 4 embedding providers; leveraged async background workers for document indexing, reducing user-perceived latency by ~90%. Implemented 3-tier JWT-based role-based access control (super_admin / admin / user) with database-per-tenant PostgreSQL isolation and Fernet-encrypted credential storage, securing 15+ API route groups across the platform. Developed a LangGraph-based ReAct custom agent builder with two-phase tool pre-selection and thread-safe credential injection via contextvars, enabling 15+ MCP tool integrations including Gmail, Slack, Drive, and Sheets. Integrated in-session document upload with direct PDF text extraction — bypassing full RAG indexing to cut processing overhead by ~80%, with upload caps at 2 MB to minimize noise and reduce hallucination.
As a Machine Learning Intern at AnuBrain Technologies, I was responsible for designing and implementing an AI-powered system for estimating the Direction of Arrival (DoA) of radio signals from antenna arrays—an essential task in wireless communication systems. The objective was to develop a deep learning-based alternative to traditional signal processing methods like MUSIC and ESPRIT, which often struggle with noise sensitivity and scalability. I built a hybrid model that combined Convolutional Neural Networks (CNN) for spatial feature extraction and Long Short-Term Memory (LSTM) networks for capturing temporal dependencies. This fusion enabled the model to learn from raw sequential antenna data and deliver highly accurate angle predictions. To enhance the learning of temporal patterns, I applied sinusoidal positional encoding and incorporated Multi-Head Attention with 16 attention heads and a key dimension of 128. This significantly improved feature representation and reduced RMSE to 5.9°. The model achieved an R² score of 0.95 and a Mean Absolute Error (MAE) of just 2.79°, clearly outperforming traditional approaches. To improve interpretability and performance, I extracted attention-based features and applied Minimum Redundancy Maximum Relevance (MRMR) to select the top 50 most informative features. For further optimization, I used Improved Grey Wolf Optimization (IGWO) with 30 agents over 15 epochs, which reduced the validation loss by 8%. I validated the model using 5-fold cross-validation and hyperparameter tuning with scikit-learn, achieving 96% of predictions within a ±5° margin of the true direction. The entire pipeline was built using TensorFlow, Keras, and Python, and reflects strong capabilities in model development, feature engineering, and optimization. This internship allowed me to work on a high-impact, research-grade problem, deepening my understanding of AI model architecture, attention mechanisms, and real-time signal estimation.