Built an Android AI agent that can see the phone screen, understand voice commands, and complete multi-step mobile tasks in real time. The cloud branch uses model APIs for fast autonomous screen control, while the on-device branch explores private local inference with speech-to-text, a vision-language model, and text-to-speech running directly on the phone.

Built Moon, an open-source touch-to-talk voice AI pendant. The ESP32-S3 runs ESP-IDF firmware for its AMOLED face, mic, and speaker, and streams 16 kHz mono audio to a Flutter and Kotlin Android app over a private BLE GATT service; 24 kHz speech comes back. A mouth-openness byte in each 12-byte packet header lets the face lip-sync against its own playback clock rather than BLE arrival time. Inference stays on the phone: Moonshine STT, Qwen3 GGUF on llama.cpp, and Kokoro ONNX.

Built a complete Speech-to-Text Transformer model from scratch using PyTorch, converting raw audio waveforms into text without pre-trained models. Implements convolutional downsampling, multi-head self-attention, Residual Vector Quantization (RVQ), and CTC loss for alignment-free training. Trained on the LJSpeech dataset using an A100 GPU.

Fine-tuned Qwen3-TTS-12Hz-1.7B-Base on 240 clips of my own voice (61.6 min), scored against Qwen's zero-shot baseline on a 20-clip holdout. Each 80 ms frame becomes 1 semantic plus 15 residual acoustic codes: the Talker predicts the semantic code, the Code Predictor adds detail, Code2Wav reconstructs PCM. One epoch on an RTX 5090 cut WER from 10.07% to 7.64%, but speaker similarity fell from 0.857 to 0.810 and the clone lost 2 of 3 blind listener votes.

Trained a causal MarbleNet-style VAD from scratch in a ~20-hour A100 run: 143k parameters, 0.55 MB in fp32, 1.71s causal receptive field. It labels each 10 ms frame speech or non-speech from 64 MFCCs through stacked 1D depthwise-separable conv blocks. Training data was LibriSpeech mixed with MUSAN noise and RIRS reverb; the checkpoint exports to ONNX. Validation F1 was 0.988 on a LibriSpeech dev surrogate, which the repo flags as a plumbing check, not a production eval.

Implemented a Tacotron 2 neural text-to-speech model from scratch in PyTorch. The model generates mel-spectrograms from raw text input using an encoder-decoder architecture with attention mechanisms, then converts them to audio waveforms. Trained on the LJSpeech dataset.

Built an OpenEnv-native scalable oversight gym for hidden sabotage detection, inspired by SHADE-Arena. A frozen DeepSeek-R1 attacker executes hidden side tasks inside deterministic Python sandboxes, while a Qwen2.5-1.5B LoRA monitor is trained with TRL GRPO and verifiable rewards to flag sabotage from public tool-call traces, reaching 0.893 AUROC with 0.88 recall and 0.12 FPR.

Built an OpenEnv reinforcement-learning environment where agents test a deliberately buggy REST task-management API. The environment includes 13 planted vulnerabilities mapped to OWASP API Security Top 10, seed-randomized data, deterministic bug detectors, a 5-signal reward function, and automatic OWASP-style bug bounty reports.

Implemented the complete DeepSeek V3 architecture from scratch, a 100M+ parameter transformer featuring Multi-Head Latent Attention (MLA), Mixture of Experts (MoE), and Multi-Token Prediction (MTP). Trained on the FineWeb-Edu dataset with ~2.5B tokens on an NVIDIA A100 80GB GPU.

Implemented a Large Language Model (LLM) from scratch, covering every stage from data preparation and model architecture to pretraining and fine-tuning. This project demystifies transformer-based models through hands-on code and experiments, enabling a deeper understanding of attention mechanisms and token prediction.

Developed an end-to-end MLOps project to detect malicious URLs using XGBoost. Integrated robust pipelines for data ingestion, model training, deployment, and monitoring.