Portfolio · Blog · Twitter · LinkedIn
mayankpratapsingh022@gmail.com
Most of my work so far has been broader deep learning: vision, multimodal, generative AI. The direction I'm actively towards is audio and speech, where the interesting problems aren't accuracy on a clean benchmark but everything around it: accents, background noise, latency budgets measured in milliseconds, and models that have to keep working after they leave the lab.
I care most about the unglamorous middle of the pipeline: the data, the evals, and the inference optimization that decides whether a model actually ships or just demos well. That usually means going past the API and rebuilding the piece I don't trust yet, so I know exactly where it breaks.
On Screen Real-time autonomous mobile agent that sees the phone screen, understands voice commands, and completes multi-step tasks. The on-device branch runs speech-to-text, a vision-language model, and text-to-speech locally.
Speech-to-Text Transformer from Scratch Encoder-decoder ASR built without any pretrained weights: convolutional downsampling, self-attention, residual vector quantization, and CTC loss, trained on LJSpeech.
Text-to-Speech (Tacotron 2) from Scratch Neural TTS from raw text to mel-spectrograms to waveform, encoder-decoder with attention, trained on LJSpeech.
Voice Activity Detection from Scratch (MarbleNet) Small, streaming VAD model trained from scratch with synthetic speech and noise mixing, full evaluation, ONNX export, and deployment-ready postprocessing.
SHADE-Gym OpenEnv-native oversight gym where a GRPO-trained monitor learns to catch a frozen model's hidden sabotage from tool-call traces alone (0.893 AUROC).
API Testing RL Environment OpenEnv environment where agents probe a deliberately buggy REST API for 13 OWASP-mapped vulnerabilities and file bug reports.
DeepSeek V3 from Scratch 100M+ param transformer with Multi-Head Latent Attention, MoE, and Multi-Token Prediction, trained on 2.5B tokens of FineWeb-Edu on a single A100.
LLM from Scratch GPT-2-style language model built from data prep through pretraining and fine-tuning, covering attention and token prediction end to end.
Network Security: Malicious URL Detection End-to-end MLOps pipeline detecting malicious URLs with XGBoost: ingestion, training, deployment, and monitoring.
