I'm an AI Engineer & Researcher with ~8.5 years of experience in Generative AI, speech synthesis, LLM architectures, agentic systems, and AI infrastructure.

I enjoy implementing research papers from scratch, training and optimizing models, exploring GPU-level inference, and building real-world AI systems.

• 🚀 Co-founder & CTO at Dubpro.ai (DeepSync Technologies) — building speech synthesis, multilingual AI, voice cloning, and generative audio technologies.
• 🎙️ Former AI Consultant at MyClone.is (San Francisco) — worked on real-time Voice AI, digital personas, RAG, memory, and agent orchestration.
• 📩 Email: rishikksh20@gmail.com

• LLMs & Architectures: Transformers, Hybrid/Linear Attention, SSMs, Mamba, model implementations from scratch.
• Training & Post-training: SFT, RLHF, RLVR, GRPO, model distillation, and alignment.
• Efficient Inference: Triton, CUDA, Flash Attention, KV caching, quantization, GPU memory and latency optimization.
• Voice & Generative Audio: TTS, voice cloning, vocoders, neural codecs, diffusion, flow matching, real-time speech systems.
• Agentic AI: Coding agents, MCP, multi-agent orchestration, RAG, persistent memory, tool calling, and sandboxing.

• MyClone — Real-time Voice AI digital persona platform with RAG, persistent memory, voice cloning, and persona-aware conversations. LiveKit LlamaIndex FastAPI pgvector

• Nexus Code Agent — CLI-first AI coding agent featuring MCP, sub-agents, skills, tools, session memory, context management, and sandboxed execution. AI Agents MCP Multi-Agent

• Voice AI Agent — Real-time conversational AI with inbound/outbound Twilio calls, browser WebRTC, and text chat. Pipecat Twilio STT LLM TTS

Minimal PyTorch implementations of modern model architectures, focusing on attention, SSMs, caching, decoding, and inference efficiency.

• LFM2.5 2.6B — Hybrid convolution/attention LLM with Triton inference kernels.
• Mamba-3 — Selective State Space Model implementation.
• Qwen3.5 0.8B — Hybrid GatedDeltaNet + GQA architecture.
• Gemma 3 270M — Lightweight LLM implementation and exploration.

Selected open-source implementations from my work in speech synthesis, neural audio, and generative modeling:

• Voxtral Codec — Neural audio codec with semantic/acoustic quantization.
• NU-Wave — Diffusion-based neural audio upsampling.
• iSTFTNet — Lightweight neural vocoder.
• VocGAN — Real-time GAN-based neural vocoder.
• AcademiCodec — Neural audio codec research.

Other paper implementations: ViViT · CvT · LocalViT · FNet · CoaT · SiT

AI & Research: PyTorch · Transformers · SSMs · RL/RLHF · Diffusion · TTS

Inference: Triton · CUDA · Flash Attention · vLLM · KV Cache · Quantization

Agents & Voice: MCP · RAG · LlamaIndex · LiveKit · Pipecat · Twilio

Backend & Infra: Python · FastAPI · PostgreSQL · Redis · Docker · AWS · GCP

Open to AI research collaborations, challenging engineering problems, and ambitious AI projects.

🐙 GitHub · 𝕏 @ai_rishikesh · 📩 Email