I'm an AI engineer and researcher with ~8.5 years of experience, working across Generative AI, speech synthesis, language models, multimodal systems, and production AI infrastructure.

I enjoy working close to the fundamentals implementing architectures from research papers, training and optimizing models, understanding inference at the kernel level, and turning those ideas into production AI systems.

I co-founded Dubpro.ai (DeepSync Technologies) and currently serve as CTO, leading AI and engineering across speech synthesis, multilingual AI, generative models, agentic systems, and AI-assisted dubbing technologies.

Alongside my work at Dubpro.ai, I recently collaborated 🔥 for six months as an AI Consultant with MyClone.is (formerly Rappo), a San Francisco-based company, where I helped build a real-time Voice AI digital persona platform capable of replicating a user's knowledge, conversational style, persona, and voice from their text and voice data.

📝 Email: rishikksh20@gmail.com

I’m always interested in taking on new technical challenges, research collaborations, and ambitious AI projects, especially where I can explore new model architectures, understand systems from first principles, and turn research ideas into practical AI products.

• Transformers, hybrid attention, linear attention, and State Space Models (SSMs)
• From-scratch PyTorch implementations of modern LLM architectures
• Model loading, KV caching, decoding, sampling, and generation pipelines

• Fine-tuning, Reinforcement Learning, RLHF, and post-training alignment
• PyTorch performance optimization, Flash Attention / SDPA and GQA
• CUDA-aware optimization and custom Triton kernels
• Optimizing token throughput, GPU memory usage, latency, and inference performance

• Text-to-Speech, voice cloning, neural vocoders, and audio codecs
• Discrete speech/audio representations
• Diffusion and Flow Matching based generative models
• Real-time STT → LLM → TTS pipelines, Voice AI agents, and telephony systems

• RAG, vector retrieval, long-term memory, and knowledge systems
• Multi-agent orchestration and agent workflows
• MCP, tools, skills, tool calling, and context/state management
• Real-time AI agents using LiveKit, Pipecat, and telephony infrastructure

Production-grade Voice AI + Digital Persona platform

An open-source version of the platform I worked on at MyClone.is.

Built around real-time voice conversations, RAG, persistent memory and persona-aware AI, with a backend involving LiveKit, FastAPI, LlamaIndex, PostgreSQL/pgvector and asynchronous data ingestion.

Voice AI · LiveKit · RAG · Memory · FastAPI · pgvector

CLI-first AI coding agent and agent runtime

A coding agent built around MCP, custom tools, skills, sub-agents, persistent sessions, context compaction, workspace memory, sandboxed execution and multi-provider LLM support.

The project is also an exploration of how coding agents can manage long-running context, permissions, delegation, tool execution and repository-level knowledge.

AI Agents · MCP · Multi-Agent · Context Management · Tools · Sandboxing

Real-time Voice AI + Telephony reference implementation

A smaller Voice AI system supporting inbound/outbound Twilio calls, browser WebRTC voice conversations and text chat using a shared real-time Pipecat pipeline.

Pipecat · Twilio · WebRTC · STT · LLM · TTS

I regularly implement recent model architectures in bare-minimum PyTorch to understand the underlying computation rather than treating models as black boxes.

• LFM2.5 2.6B — hybrid convolution/attention LLM, KV-cache optimization and custom Triton inference kernels
• Mamba-3 — clean implementation of the latest selective State Space Model architecture
• Qwen3.5 0.8B — hybrid GatedDeltaNet linear attention + GQA implementation
• Gemma 3 270M — lightweight implementation and model exploration

My focus here is generally on understanding attention, SSMs, model state, caching, decoding, memory movement and inference performance.

Speech synthesis has been one of my longest-running research areas.

Some selected implementations:

• Voxtral Codec — neural audio codec with semantic/acoustic quantization
• NU-Wave — diffusion model for neural audio upsampling
• iSTFTNet — lightweight neural vocoder using inverse STFT
• VocGAN — real-time neural vocoder and modified adversarial training architecture
• AcademiCodec — open-source neural audio codec research

My broader work includes TTS, voice cloning, vocoders, audio codecs, diffusion models, flow matching and real-time speech generation systems.

I also maintain implementations of research papers to study architectures from first principles:

ViViT · CvT · LocalViT · FNet · CoaT · SiT

These projects span transformers, convolution-attention hybrids, Fourier token mixing, self-supervised learning, video transformers and generative modeling.

Models & Research: PyTorch · Transformers · SSMs · Diffusion · Flow Matching · RL/RLHF · TTS

Inference: Triton · CUDA-aware PyTorch · Flash Attention / SDPA · vLLM · KV Cache · Quantization

Agents: MCP · RAG · LlamaIndex · LangChain · Multi-Agent Systems · Tool Calling · Long-term Memory

Voice: LiveKit · Pipecat · Twilio · Deepgram · AssemblyAI · ElevenLabs · Cartesia

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

I'm particularly interested in AI research, LLM systems, Voice AI, intelligent agents, model architecture and efficient inference.

• GitHub: github.com/rishikksh20
• X / Twitter: @ai_rishikesh
• Email: rishikksh20@gmail.com