Production Speech-to-Speech pipelines, distributed NATS messaging, and physical kinematics engineered at Machani Robotics.
I am Kavin B M, an AI Engineer based in Chennai and Bangalore. I build the engineering around models: sub-200ms discrete audio tokenization, 10+ MCP server fleets over NATS, and safe 12-DOF physical motor actuation.
Neural audio codecs with edge wakeword engines.
Synchronizing LLM reasoning with motor actuation.
10+ custom MCP servers orchestrated over NATS.
Multi-layer prompt injection defense and evals.
Voice interactions break down when latency exceeds human turn-taking thresholds. I build discrete speech pipelines that tokenize audio directly at the edge, streaming neural representations to avoid round-trip cloud bottlenecks.
Edge wakeword engines operate locally for zero audio leakage. Real-time prosody analysis and emotion conditioning let the system interrupt naturally, respect user cadence, and respond with human conversational timing.
Explore Robotics ActuationOrchestrating isolated tool runtimes with low-latency pub/sub messaging, typed gRPC contracts, and deterministic tool pruning.
Sub-millisecond pub/sub event mesh enabling instant tool discovery and telemetry streaming.
<1msSub-1ms fast-path exact lexical match bypassing vector embeddings and LLM reranking.
Go / BM25Reactive desktop control plane for live server state inspection, session debugging, and dispatch.
Rust / UIStrict Protobuf contracts translating LLM tool calls into physical robotics joint commands.
ProtobufRefusing to route low-confidence inputs to avoid errant tool mutations on physical hardware.
PolicyAI Engineer · Bangalore, India (Jan 2026 - Present)
Production Speech-to-Speech (S2S) voice pipelines with sub-200ms TTFT, 12-DOF embodied servo kinematics, 10+ custom MCP servers over NATS, and OpenTelemetry observability.
College of Engineering, Guindy (Anna University)
2021 – 2025 · CGPA 8.0/10
Go · Python · TypeScript · Rust · SQL
NATS · MCP · gRPC · Protobuf · Tauri · Docker