PlayStation Trust & Safety LLM Toxicity Rewriting

• Developed a Generative AI and NLP system for Sony PlayStation that rewrites toxic multiplayer messages into neutral language while preserving semantic intent.

• Built the GPT-2 Medium decoder-only Transformer in PyTorch, initialized it with OpenAI pretrained weights, and performed supervised fine-tuning and instruction tuning on ParaDetox using response-only causal language-model loss.

• Engineered the end-to-end LLM pipeline using Hugging Face Datasets, tiktoken BPE tokenization, dynamic batching, AdamW optimization, gradient clipping, checkpointing, and autoregressive inference.

• Achieved approximately 90% toxicity removal, 0.86 semantic similarity, 0.88 fluency, and sub-300 ms GPU latency, with competitive performance against BART, CondBERT, and ParaGeDi baselines.

Vision Transformer-Based Wildlife Autofocus System for SONY Alpha 7S camera

• Built an end-to-end wildlife autofocus system that uses a GPT-2-style Vision Transformer to detect animals, preserve the selected target across video frames, and generate continuous lens-focus coordinates, achieving ~56 [email protected]:0.95 and around 90% target retention.

• Improved autofocus stability by combining object detection, temporal association, and target-selection logic within the same inference pipeline, reducing unintended focus switches by ~30% during rapid motion, partial occlusion, and foliage interference.

• Optimized the complete PyTorch computer-vision pipeline for embedded camera deployment using INT8 quantization and inference-graph optimization, reducing model footprint by ~75% while sustaining ~22 FPS with less than a 2-point mAP reduction.

• Developed deep learning-based image restoration solutions using U-Net architectures in TensorFlow for image denoising and deblurring applications, improving PSNR from ~26 dB to ~31-32 dB and SSIM from ~0.45 to ~0.85-0.90 over classical filtering methods.

• Built scalable data preprocessing, augmentation and dataset preparation pipelines for paired noisy-clean image datasets used in supervised learning workflows

• Performed systematic hyperparameter tuning and model refinement, improving image quality metrics including PSNR (~26 dB → ~31-32 dB) and SSIM (~0.45 → ~0.85-0.90), while classical methods remained effective on only a single noise type.

• Conducted performance evaluation and comparative benchmarking to identify optimal model configurations for deployment, confirming consistent generalization across noise types where classical baselines failed