Narain Pattabhiraman

Built an agent that interprets natural-language specs to autonomously write, execute, and debug code, then render interactive UI widgets (charts, forms, dashboards) as live browser components.

Integrated computer-use capabilities — screenshot perception and DOM interaction — enabling the agent to visually verify and self-correct rendered output in a sandboxed browser environment.

Orchestrated multi-step tool chains (code generation, execution, visual feedback, iterative refinement) using a ReAct loop with MCP-compatible tool adapters.

Built a multi-stage synthetic data generation pipeline producing JSON-formatted reasoning QA pairs via structured relationship sampling, entity assignment, and bidirectional validation.

Fine-tuned Qwen3 14B with LoRA + SFT, then post-trained with GRPO using weighted multi-component rewards (correctness, format, completeness) to improve reasoning-chain quality.

Designed a Unity3D simulation environment in PyTorch and built a physical prototype with 3D-printed components.

Trained deep RL agents (DDPG, A2C, PPO) for continuous control tasks with >30 DoF using neural policies over sensor and image inputs, then transferred simulation-trained policies to real-world hardware.

Built zero-install browser inference for Gemma, LLaMA, and Mistral using WebAssembly SIMD, enabling local LLM execution without backend infrastructure.

Adapted LLaMA for multilingual and domain-specific use cases by extending the tokenizer and training on translated instruction datasets and domain-specific dialogue corpora.

Engineered a graph-optimization compiler for mobile inference backends using TVM and TASO; wrote ARM NEON SIMD and custom CUDA kernels for Jetson Orin with measurable speedups over OpenCV C++ and TorchScript.

Built a CNN-based steganalysis model for detecting hidden content in digital images by modifying ResNet with Efficient Channel Attention and reduced convolution stride for improved sensitivity.