Rohith K Bobby
Mavelikara, Kerala, India
Machine learning & backend developer, voice AI wrangler, bug magnet, and occasional genius.
About Me
I build end-to-end ML systems that move from model experiments into reliable production services. My current work focuses on low-latency voice AI pipelines, self-hosted ASR and TTS, Ray Serve, Kubernetes, and GPU-backed inference that can handle real traffic.
Core Toolkit
Experience & Education
ML & MLOps Engineer
12/2025 – PresentGenerativeStudio.dev · Remote
- Built low-latency voice AI pipelines with LiveKit, self-hosted ASR, LLM, and TTS services for production call handling.
- Served Parakeet ASR and Kokoro TTS on bare-metal L40S infrastructure with Kubernetes, KubeRay, Ray Serve, and Docker.
- Validated inference services through concurrency sweeps, reaching 155 RPS for ASR, 56 RPS for TTS, 630x real-time audio throughput, p50 latency below 200ms, and full transcription match across 2,400 test requests.
- Reduced cold-start latency and timeout failures in async Ray Serve deployments by tuning worker concurrency, batching, and deployment settings.
Machine Learning Intern
03/2023 – 12/2023IHRD / SBCID Kerala
- Built an automated pipeline for OCR and news article summarization, reducing processing time by 75%.
- Implemented NLP-based keyword extraction to tag articles and improve case identification accuracy by 20%.
- Developed a text summarization and translation pipeline for Indic languages, improving BLEU scores by 25% and used by 20+ officers.
B.Tech in Computer Science, Cyber Security
11/2022 – 04/2026College of Engineering Kallooppara
Completed B.Tech in Computer Science with a Cyber Security focus. Dissertation covered fast fully homomorphic encryption via Module-LWE with parallel NTT algorithms, GPU Roofline modeling, and IND-CPA security analysis.
Projects
TinyServe
- Building a minimal local LLM inference server for developers who need simple model serving without the overhead of larger serving stacks.
- Added OpenAI-compatible chat completions, streaming responses, request queuing, token usage logs, per-model config, GPU memory display, request cancellation, and model warmup.
Offline Voice Assistant for Field Technicians
- Led ML architecture for a fully offline edge AI voice assistant at the Armada Hackathon, built for field technicians without internet access.
- Combined local ASR, noise suppression, Vision-RAG document retrieval, vision-language reasoning, and TTS into one offline assistant.
goKyber
- Implemented the Kyber key encapsulation mechanism in Go with clean structure, practical benchmarking, and educational clarity.
- Compared lattice-based key exchange against RSA and ECC baselines to study performance tradeoffs.
Get In Touch
Open to ML infrastructure, voice AI, backend, and applied research work. Feel free to reach out.