"I don't play the odds. I play the data."

I am a Data Science Student at IIT Madras, focused on pushing the boundaries of machine learning and artificial intelligence. My work heavily revolves around deep learning architectures, data-driven solutions, and contributing to the core infrastructure of the biggest ML frameworks in the world.

• I am currently working on advanced projects involving TensorFlow, Keras, and Generative AI.
• Exploring the bleeding edge of Deep Learning and neural networks.
• Ask me about: Python, C++, Machine Learning, GPU Compilers, and AI models.

I do not just use open-source tools; I build and fix their core engines. I actively contribute to ecosystems like Google, LangChain (langchain-ai), Pydantic, and AutoGPT. Here are some of my top Pull Requests structured using the STAR (Situation, Task, Action, Result) methodology:

• Google MCP Toolbox - PR #3506 & #3660
  • Situation/Task: Enterprise data pipelines required strict compliance for data security and high-throughput serialization without latency bottlenecks.
  • Action: Engineered policy-based automatic PII detection and optimized row serialization and JSON parsing mechanics.
  • Result: Secured data handling and increased throughput efficiency for enterprise-scale data pipelines.
• LangChain - PR #1590
  • Situation/Task: ChatAnthropicVertex integrations lacked model-aware token constraints, risking truncation errors during execution.
  • Action: Developed dynamic, model-aware max_output_tokens management tailored for Google Vertex AI integrations.
  • Result: Prevented generation truncation errors, ensuring stable, production-ready LLM orchestration.
• Pydantic AI - PR #4276
  • Situation/Task: AI validation pipelines were prone to failing due to stale client references remaining in model and embedding classes.
  • Action: Refactored state-management logic to ensure stale references are purged and properly re-initialized.
  • Result: Restored lifecycle stability and reliable execution across AI validation workflows.
• AutoGPT - PR #12089
  • Situation/Task: Autonomous agent memory systems crashed in constrained database environments where pgvector types were unavailable.
  • Action: Built a graceful degradation architecture for vector search fallbacks.
  • Result: Drastically improved the resilience and uptime of autonomous agents in non-vectorized environments.
• TensorFlow Core (Merged) - PR #123489 & #123491
  • Situation/Task: Data loading pipelines in TensorFlow's C++ core suffered from silent failures and performance degradation.
  • Action: Fixed an unguarded vector resize and patched a critical signed-to-unsigned conversion bug in IteratorRandomAccessCache::Get.
  • Result: Stabilized C++ core memory caching, preventing data corruption for global AI infrastructure.
• Keras - PR #23552
  • Situation/Task: The mathematical core of the categorical_crossentropy loss function contained a zero-day vulnerability.
  • Action: Patched the underlying logic to completely prevent division-by-zero NaN outputs.
  • Result: Ensured stable, mathematical backpropagation for classification models worldwide.
• OpenXLA Compiler - PR #47978
  • Situation/Task: Next-gen Nvidia Blackwell (sm_120) architectures lacked proper instruction set minimums in the NVPTX backend.
  • Action: Configured and forced the PTX 8.7 minimum instruction set directly within the compiler backend.
  • Result: Guaranteed hardware compatibility and execution optimization for state-of-the-art AI accelerators.
• Ubuntu Desktop (Merged) - PR #1510
  • Situation/Task: The Ubuntu OS installer suffered a fatal UI crash when users selected Right-to-Left (RTL) locales.
  • Action: Engineered a fix forcing LTR directionality on critical installer UI components.
  • Result: Restored seamless OS provisioning and accessibility for global RTL language users.

I design solutions for everything from enterprise cloud architecture to aerospace optics. Here is what I am building:

• Project Radius
  • Focus: High-Performance Adaptive Optics C-Engine for Shack-Hartmann Wavefront Sensing (ISRO Compliant).
  • Impact: Aerospace-grade optics engineering. Demonstrates the intersection of advanced physics, sensor processing, and high-performance computing.
• CoT Rationalization Forensics
  • Focus: Chain-of-Thought (CoT) alignment and rationalization for Large Language Models.
  • Impact: Pushing the boundaries of AI interpretability, reasoning tracing, and alignment in modern generative models.
• Spanner Vertex Search
  • Focus: Enterprise search architecture utilizing Google Cloud Spanner and Vertex AI.
  • Impact: Highly scalable, AI-driven vector search implementation leveraging enterprise cloud infrastructure.

My Contribution Graph (Eating the data, one commit at a time):