KBM.
BLR · AI Engineer @ Machani Robotics

00 — Engineering Portfolio

Kavin B M

Applied AI & Robotics Systems Architecture

Building production AI infrastructure, real-time Speech-to-Speech voice pipelines, and embodied robotics actuation.

Explore Flagships↘Résumé PDF↗Quick Transmission ✉
180ms TTFTDiscrete S2S Audio<1ms PathMCP Dynamic Router10+ DaemonsNATS Tool Mesh12-DOF LoopServo Kinematics
Scroll down to inspect live systems pipeline↓

00.1 — Architecture Blueprint

Live Speech-to-Speech & Robotics Dataflow

Production audio, NATS messaging, and motor kinematics running at Machani Robotics.

Stage 02//Discrete Neural Tokens

Neural Codec & S2S

Direct discrete Speech-to-Speech pipeline. Quantized audio tokens map speech directly into model reasoning space and back to streaming vocoder synthesis without cascading STT/TTS latency.

Latency:180ms streaming TTFT
Codec:8 Discrete Codebooks
Turn-Taking:Proactive acoustic VAD
Output:Streaming neural vocoder
// s2s.spec.tsBudget: 180ms TTFT
# Discrete acoustic tokenization to streaming synthesis
async for audio_chunk in mic_stream:
    acoustic_tokens = codec.encode(audio_chunk)
    async for gen_token in s2s_model.generate_stream(acoustic_tokens):
        raw_pcm = vocoder.decode_chunk(gen_token)
        await audio_output.write(raw_pcm)

01 — Selected Systems

Production systems built for speed, safety, and scale.

Not demo wrappers. These handle sub-millisecond routing, multi-parser document screening, distributed NATS messaging, and physical robotics actuation.

Flagship Case Studies (01 — 03)Deep Architectural Dives
// 01Open source

ResumeShield

Finds prompt injection hidden inside resumes

Problem //AI resume screeners read text a human reviewer never sees, so an instruction hidden in white-on-white 1pt type can change a hiring decision.

Decision //Detect the property rather than the phrase — layered structural, parser-divergence, encoding and behavioural checks, so no single rewording evades them.

PythonFastAPIPyMuPDFNext.jsLiteLLM
Proven OutcomeHidden instructions caught and stripped before screening
Inspect↗
// 02Open source · ★ 7

MCP Dynamic Router

Real-time tool selection for conversational AI

Problem //Tool-heavy agents get slow and unpredictable when every option reaches the model.

Decision //Narrow the search space first, then rank the remaining tools with a hybrid routing pipeline.

GoMCPBM25OllamaRAG
Proven Outcome<1ms fast path for exact matches
Inspect↗
// 03Production Tool

MCP Mesh & Desktop Hub

Fleet control plane for 10+ MCP servers over NATS

Problem //Coordinating 10+ MCP servers across disconnected runtimes introduces IPC bottlenecks and fragmented tool state.

Decision //Pair a lightweight Tauri + Svelte desktop UI with a NATS pub/sub message bus for instant, bi-directional tool communication.

TauriSvelteNATSGoMCPRust
Proven Outcome10+ MCP servers orchestrated via low-latency NATS mesh
Inspect↗
Systems & Telemetry Fleet (04 — 07)Go · Python · OpenTelemetry · gRPC
04Open source

GhostInt

AI OSINT · digital footprints as a living 3D graph

Next.js 14Three.jsBun
Multi-source intelligence as a navigable 3D graph↗
05Open source

Burnmeter

Local-first LLM cost observability

PythonSvelteOpenAI API
Private, on-device usage and spend visibility↗
06Open source

Spotify gRPC-MCP

Spotify as AI-native tools

GogRPCMCP
Spotify controls exposed as agent-ready tools↗
07Contribution

Pulse Observability

Unified telemetry framework

GoOpenTelemetry
Clearer operational signals across services↗
Interactive Lab [WebGL]08 // Spatial Architecture

AI Systems 3D Architecture Lab

Walk through an interactive 60 FPS 3D spatial breakdown covering discrete Speech-to-Speech audio pipelines, NATS tool orchestration, and physical servo kinematics.

Launch 3D Lab↗

02 — How I think

Three things I believe about building this.

I care about what happens after the demo: whether it's fast enough, whether it fails safely, and whether you can tell what went wrong. Most of my work is the engineering around a model rather than the model itself.

01

Latency is a feature, not a footnote.

Users forgive a model that is occasionally wrong. They do not forgive one that is slow to act. So I design the fast path first and treat everything else as the exception.

02

A system should know when to say nothing.

Most failures in AI tooling are not wrong answers, they are confident ones. Abstention is a design decision, and it belongs in the architecture rather than in a prompt.

03

If you cannot see it, you cannot ship it.

Anything running in production needs to be measurable by the person responsible for it, without sending that data somewhere else first.

03 — Path

Work that taught me to ship.

2024→NOWRésumé ↗
01Jan 2026 - Present

AI Engineer

Machani Robotics · Bangalore, India

  • Production Speech-to-Speech (S2S) pipeline: low-latency streaming audio architecture with edge wakeword detection, proactive turn-taking, and emotion-conditioned voice synthesis.
  • Local discrete S2S & embodied actuation: researched on-device neural audio tokenization and low-latency speech pipelines synchronized with motor/servo kinematics for expressive physical robot interaction.
  • MCP Tooling & Messaging Mesh: architected and orchestrated 10+ custom MCP servers interconnected via NATS message broker with a reactive Tauri + Svelte desktop control interface.
  • Hybrid gRPC-MCP architectures: designed typed low-latency IPC bridging multi-agent LLM reasoning with robotics actuators.
  • Observability & evaluation: deployed OpenTelemetry distributed tracing and real-time telemetry dashboards for inference latency, token costs, and safety abstentions.
02Jan 2025 - Jun 2025

AI Developer Intern

Brixbit Technologies · Chennai, India

  • End-to-end AI workflows with n8n and Replit for rapid prototype deployment.
  • RAG applications using LangChain and LlamaIndex.
  • Agentic workflows for contextual retrieval and dynamic response generation.
03Jun 2024 - Jul 2024

Software Engineer Intern

PreludeSys India · Chennai, India

  • Design, development, testing and debugging across the software lifecycle.

+ Academic Foundation

2021 — 2025

B.E. Computer Science & Engineering

College of Engineering, Guindy (Anna University)

CGPA 8.0/10 · Focus on Distributed Systems, OS, Compilers, and AI Architecture.

2019 — 2021

Higher Secondary (CBSE)

Chennai Public School

Computer Science & Mathematics · 93%

⚡ Technical Stack Matrix

Languages
GoPythonTypeScriptRustSQL
AI Systems & Audio
Discrete S2S PipelinesNeural Audio CodecsWakeword EnginesProactive AgentsLLMs & RAGModel EvaluationPyTorch
Robotics & Protocols
Motor & Servo ActuationMCPNATSgRPCProtobufWebSocketsA2AREST
Infra & Desktop
TauriSvelteOpenTelemetryRedisPostgreSQLDockerDistributed Tracing

04 — Signal & Dispatch

Got something hard to build?
Let's connect.

Bangalore / ChennaiIST

1. Pick your conversation vibe:

2. Auto-generated dispatch payload:Let's bring you onto our team to build high-impact AI systems.
SUBJECT: Engineering Opportunity: Applied AI / AI InfrastructureTO: kavinbm16@gmail.com

Hi Kavin, I was exploring your work on MCP fleets, local discrete S2S pipelines, and robotics infrastructure. We have an engineering role that would be an exceptional fit for your background. Let's set up a time to chat!

🚀
Transmit SignalLaunch Email Client with Payload ↗
Send ↗
GitHub↗
@Kavin-bm
10+ repos · MCP Router · GhostInt · S2S
LinkedIn↗
in/kavinbm
Connect, chat, & engineering network
Résumé↗
PDF ↗
Production metrics, systems & CV