AI Engineering & Research Studio

Disciplined orchestration
beats bigger models

We make AI boring, dependable, inspectable, and production-worthy. Murai Labs is an embedded research studio for teams building AI that needs to actually work.

Start a Conversation See Our Research
Our Philosophyமுறை — method, order, disciplined process
Most AI teams chase parameter counts. We chase reliability. Our work sits at the intersection of rigorous research and production engineering — we build systems where every decision is inspectable, every output is traceable, and every failure mode is anticipated. Disciplined orchestration isn't a constraint. It's the architecture.
What We Doகலவை — fusion, blend
We work as an embedded AI research team — from architecture design through validated deployment. Three ways to work with us.
01

Model Fusion & Unification

Your team has fine-tuned multiple specialist models that sit in silos. We fuse them into a single, routable endpoint — validated across domains, with deterministic dispatch and measurable per-domain performance. No retraining required.

KALAVAI Protocol
02

Embedded AI Research

A senior research engineering partner embedded in your team. Architecture reviews, multi-agent system design, training strategy, evaluation rigor. We bring the discipline of published research to your production timeline.

Retainer
03

On-Device & Edge AI

Inference on Apple Neural Engine, local-first architectures, cloud-fallback routing. We've published the first open system for programming Apple's ANE directly — and we bring that depth to your edge deployment challenges.

Orion Framework
Resultsவிளைவு — outcome
+21.76%

Cross-Lingual Fusion Gain

Tamil, Yoruba, Welsh, and Code specialists fused into a single model with validated held-out improvement across all domains.

+10.17%

Private Domain Fusion

Medical, legal, and patent specialists unified with 3-seed validation. Conversion rate of 0.55× divergence — consistent and reproducible.

8.5×

ANE Compilation Speedup

Delta compilation for Apple Neural Engine inference — first open system bypassing CoreML for direct ANE access.

99.8%

Router Determinism

MoE dispatch routing across all KALAVAI configurations. Each specialist matched on its domain with near-perfect routing accuracy.

Researchஆய்வு — inquiry

KALAVAI — Cooperative LLM Fusion Protocol

Twenty strangers train one slice of a model each. A divergence law predicts exactly how much better the fusion will be. MoE routing does the rest.
Read More →

Orion — Programming Apple's Neural Engine

The first open system that talks directly to Apple's ANE silicon — bypassing CoreML to unlock full hardware throughput for LLM inference and training.
Read Paper →

ACAR — Adaptive Complexity Routing

Routes each query to the right model at the right cost. An auditable decision trace shows exactly why. Multi-model ensembles without the inference tax.
Read Paper →

The Coherence Trap

How multi-agent systems silently collapse into consensus — and the four measurable signals that warn you before they do.
NeurIPS 2026

ĀTAVI — Multi-Agent Research Protocol

A host-agnostic agent orchestration layer that scales team size dynamically and actively resists coherence collapse.
View on GitHub →
Productsஉற்பத்தி — creation, output
Shipped products, hackathon builds, and upcoming launches — AI that exists outside the lab.

StoryForge AI

Live, production AI storytelling for kids. Children co-create illustrated adventures with an AI narrator that adapts to their age, vocabulary, and imagination in real time.

Live storyforgekidsai.com →

Vithai.AI

Voice-first bilingual memory capture for diaspora families. Grandparents speak in their language. The stories survive, translated and preserved, for the next generation.

Coming Soon vithai.ai

Blueflame

Microsoft AI Dev Days Hackathon submission. An AI-native development accelerator — turning rough intent into structured, executable engineering plans at speed.

Shipped GitHub →

Katha

AI Collective Data Portability Hackathon submission. Your data, your story — a portable AI identity layer that travels with you across platforms without re-training from scratch.

Shipped GitHub →
Founderநிறுவனர்
Ramchand Kumaresan

Ramchand Kumaresan

Founder & CEO, Murai Labs

AI research engineer with a program management background (PgMP) and deep experience shipping production software at Procore Technologies, where he worked on construction SaaS at scale — field operations, data platforms, and marketplace integrations.

Now building at the intersection of rigorous ML research and production engineering. Author of Orion (arXiv 2603.06728), the first open system for programming Apple's Neural Engine directly. Leading KALAVAI, a cooperative LLM fusion protocol targeting NeurIPS 2026. Research interests span multi-agent orchestration, on-device inference, and the governance of AI systems — grounded in classical frameworks from Tamil Sangam literature and the Arthashastra.

Based in Texas. Ships code, writes papers, builds systems that work.

PgMP arXiv Published NeurIPS 2026 Ex-Procore Apple ANE Multi-Agent Systems
LinkedIn GitHub arXiv

Let's make your AI
production-worthy

Whether you need specialists fused, agents orchestrated, or a research partner embedded in your team — start with a conversation.

Work With Us
ramchand@murailabs.com