Skymel recently unveiled its Orchestrator Agent (OA) – the industry’s first real-time AI deployment agent – designed to adaptively manage where and how AI models run across both local devices and the cloud. Built on the company’s core NeuroSplit technology, the Orchestrator Agent allows developers to configure infrastructure decisions instantly, eliminating long deployment cycles. Pulse 2.0 spoke with Skymel co-founders Neetu Pathak (CEO) and Sushant Tripathy (CTO) to learn more about their backgrounds, the motivation behind the company, and how their hybrid AI infrastructure is changing the way intelligent applications are built and delivered.
Pulse 2.0 (Amit): Could you tell me more about your background?
Skymel (Neetu): I co-founded Skymel following roles in product and engineering at Redis and Fortella, and hold a Master’s degree in Computer Science from the University of Utah. I now lead the company as CEO alongside my co-founder, Sushant Tripathy, Skymel’s CTO. Sushant previously led machine learning initiatives at Google and PayPal, holds a PhD in Computational Biology from Northwestern University, and earned his undergraduate degree from the Indian Institute of Technology, Kharagpur.
Pulse 2.0 (Amit): How did the idea for the company come together?
Skymel (Sushant): While working on on-device ML for Google Assistant, I realized how much underused compute power exists on personal devices such as smartphones and laptops. This underutilization is a hallmark of the cloud-centric services that are prevalent at this time. In order to provide personalized services, most companies prefer to upload raw user data rather than process it directly on the device.
Skymel (Neetu): At the same time, I observed a few emerging trends. Many companies were willing to pay a premium to reduce response latency in their user-facing applications. However, most lacked the in-house expertise, or had poor communication among teams, which led to inefficient use of cloud resources and escalating infrastructure costs.
Recognizing these challenges, we built NeuroSplit, our hybrid inference technology that splits AI workloads between local devices and the cloud. For enterprises, NeuroSplit delivers lower response latencies while cutting AI inference costs on existing infrastructure by up to 60%.
We then extended NeuroSplit into the Orchestrator Agent to help enterprises manage AI deployment with greater flexibility and intelligence. The OA builds on NeuroSplit’s foundation, allowing teams to adapt their infrastructure to changing needs without the complexity of constant reengineering.
Pulse 2.0 (Amit): What makes your core technology unique?
Skymel (Neetu): We are one of the first few companies to offer developers access to Hybrid AI Infrastructure (combining Local + Cloud), with a unique capability to split individual inferences. Our commercial platform empowers enterprises to seamlessly integrate local computing on end-user devices with backend cloud or on-premises servers, optimizing the performance of AI-powered applications.
Skymel (Sushant): Our NeuroSplit solution dynamically evaluates available idle compute on the end-user’s device, prioritizing it to run the AI application while utilizing cloud compute only when needed to ensure optimal performance. It is specifically engineered to off-load AI compute tasks onto end-users’ devices, based on the alignment between the device’s available idle compute, and the compute requirement of appropriate stub models that partially process the user’s data. NeuroSplit subsequently routes the stub model’s outputs to corresponding cloud model endpoints to complete the AI inference tasks. This allows NeuroSplit to scale more efficiently than pure cloud offerings, as it distributes the inference load between backend cloud servers and end-user devices. As a result, it helps AI-powered application providers reduce the specifications of backend GPU servers (both in VRAM and compute frequency) and eliminates the need for excessive standby over-provisioning (typically around 30%to handle traffic spikes) of their AI servers.
NeuroSplit reduces the amount of cloud-compute needed by AI powered applications by 30-40%, lowering the memory and compute requirements for cloud GPUs. This enables the use of lower-cost older-gen GPUs, without compromising the user experience. As a result, enterprises can achieve significant cost savings, since older-generation GPUs are more widely available and less expensive. For example, an application that once required multiple Nvidia A100s at an average cost of $2.74 per hour can now use multiple V100/Ss at 83 cents per hour with NeuroSplit, all while maintaining or even improving response latencies expected by users.
Skymel (Neetu): Enterprises running applications on end-user devices are highly motivated to provide a highly engaging user experience while keeping their cloud costs low. NeuroSplit caters to both of these needs without requiring a complete codebase overhaul or changes to cloud infrastructure. It can be easily integrated into client-facing applications with just a few lines of code.
Pulse 2.0 (Amit): You recently launched a new product. What is the Orchestrator Agent?
Skymel (Neetu): The Orchestrator Agent began as a lightweight deployment layer, allowing teams to go from idea to deployed AI feature without building bespoke infrastructure. Developers specify high-level intent (e.g., “summarize”, “transcribe”, “classify”) and constraints (like latency or cost), and the OA dynamically configures the execution environment.
Skymel (Sushant): As we engaged more with early partners, we saw a new need emerge: pipeline-level reasoning. For example, the choice of a summarization model might affect how an extraction model behaves downstream. This meant static chaining wouldn’t cut it – what they really needed was dynamic, state-aware model selection.
Skymel (Neetu): Today, most AI features – whether it’s a portfolio assistant or a document Q&A tool – are built as hardcoded pipelines. Developers wire together steps like transcription → parsing → LLM → post-processing by hand. These pipelines are brittle, expensive to maintain, and hard to adapt as user needs or models evolve. Every small change – a new input format, a stricter latency target, or a different model – requires a rewrite.
Skymel (Sushant): Developers also end up making painful trade-offs. To hit a 3-second latency target, they might downsize to a weaker model. To save cost, they might remove a key validation step. But all of this is being done by hand, and the burden of managing latency, accuracy, fallback logic, and cost – all while keeping the full pipeline functional – grows fast.
That’s what the Orchestrator Agent solves. Developers tell OA the purpose of the AI feature, like:
“This is a portfolio assistant that helps answer regulatory questions, but doesn’t give investment advice. It must respond in <4 seconds, cost <$0.05, and support both Excel and voice input.”
From there, the OA takes over. When the user writes a prompt, it:
- Breaks it into subtasks
- Chooses the right model/tool per step
- Builds the entire workflow graph in real time
- Executes the pipeline end-to-end, while optimizing for latency, cost, and correctness, not just per step, but for the whole task
Most “model routers” optimize locally – they pick the cheapest or fastest model at each step. But this misses the forest for the trees. A step might save 200ms, but break the output formatting needed for the final summarizer. Or a cheaper model might hallucinate something that later tools just repeat, not fix.
Skymel avoids this trap by reasoning over the full pipeline. It sees how upstream choices affect downstream quality, and adapts dynamically to deliver the best overall result, not just the cheapest inference.
Skymel (Neetu): To give you a concrete example, imagine a prompt like:
“Here’s my latest sales call transcript.
Extract key objections, compare them with last quarter’s trends, and generate three slides to present at next week’s meeting.”
OA might build a pipeline like:
Transcriber (Whisper) → Objection Extractor (custom prompt over Claude) → Trend Lookup (internal API) → GPT‑4o → Slide Generator
This is assembled on the fly – tailored to the prompt, the developer’s constraints, and the tools/models available.
And if latency becomes tight or privacy matters (e.g., for healthcare or finance), OA can optionally run some steps on-device using our NeuroSplit engine. But the focus isn’t on hybrid execution for its own sake – it’s on delivering adaptive, end-to-end AI pipelines that actually work.
Pulse 2.0 (Amit): What are the benefits of using the Orchestrator Agent?
Skymel (Neetu): One major benefit is speed: with OA, you can go from idea to production in days, not weeks. You define the feature intent, constraints like latency or cost, and any usage boundaries, and OA dynamically handles the orchestration. There’s no need to re-architect pipelines every time a new model drops or a new tool gets added to your stack.
Another key advantage is progressive adoption. We know not every team is ready to give up full control of their orchestration logic. That’s why the OA is designed to plug in as a dynamic subroutine inside your existing system. You can start small – offload just one step, like summarization or evaluation, to the Orchestrator Agent. Over time, as trust builds, you can expand its role. You don’t need to rip out what already works to get value from Skymel.
Skymel (Sushant): Many tools today route each step to the cheapest or fastest model. But that often breaks the full task – saving 100ms early on might ruin the quality downstream. Skymel doesn’t just optimize step-by-step. It considers the entire pipeline and makes choices that improve the overall outcome, not just the parts. You get better final results without compromising control.
It also provides deep observability. Every decision – model selection, fallback path, latency tradeoff – is logged. This creates a real-time, production-grade understanding of how your AI features behave. You can improve outputs without redeploying, and run A/B tests or error analyses instantly.
Skymel (Neetu): And perhaps one of the most exciting benefits: we’re empowering a new generation of builders.
There’s an entire wave of vibe coders – developers prototyping fast, remixing tools, shipping ideas on Replit, Cursor, or LangChain. They’re powerful in spirit but often blocked when it comes to shipping complete, robust, complex AI features – things like chaining APIs, validating responses, calling tools, or adapting workflows at runtime.
With OA, they just prompt. They describe what the feature should do, and we dynamically generate the pipeline that actually makes it work. No backend surgery. No orchestration tangle. Just real, complex AI features, integrated into their app in minutes.
In short, the Orchestrator Agent gives you adaptability, observability, and real-time orchestration – without sacrificing control. Whether you’re a platform team or a solo builder, it helps you ship stronger AI features faster, and with less fragility.
Pulse 2.0 (Amit): What has been a standout moment for Skymel so far?
Skymel (Neetu): Recent standout moment was when we saw the first real-time, multi-step AI workflow get generated and executed completely by the Orchestrator Agent – no glue code, no scaffolding, just a clean pipeline built on the fly from a user prompt. It was the first time we saw the system go beyond static orchestration and actually reason through how to structure a full task dynamically, with model selection, tool chaining, and execution – all stitched together in real time. That was the moment we knew we weren’t just optimizing infrastructure – we were reshaping how AI features get built.
Pulse 2.0 (Amit): Have you faced any challenges in your sector of work recently?
Skymel (Neetu): One challenge has been timing. When we first introduced NeuroSplit, our focus was on performance and cost optimization – splitting inference between cloud and end-user device. But most companies weren’t even there yet. They were still trying to figure out how to use LLMs in their products, not how to optimize production workflows.
So we shifted. We realized the real blocker wasn’t just infrastructure – it was orchestration. Teams were stuck building brittle pipelines by hand, and the second a new tool, model, or user request came along, things broke. That’s what led us to create the Orchestrator Agent. But like any new category, it took work to educate people that you don’t need to choose between control and flexibility – you can have both.
We’ve also had to overcome the assumption that AI infrastructure is just about model hosting or latency. In reality, the bigger problem is feature velocity – how quickly teams can deliver useful, adaptive AI experiences. That’s where we’re focused now.
Pulse 2.0 (Amit): What have been some of the company’s most significant milestones?
Skymel (Neetu): One of our first major milestones was launching NeuroSplit, our hybrid inference technology. It was a technically ambitious idea – splitting AI model execution across local devices and the cloud to reduce latency and cost – and it laid the foundation for everything that came next.
From there, we quickly realized that while hybrid execution solved infrastructure pain, the deeper problem was workflow orchestration. That realization led us to build the Orchestrator Agent (OA), and with it, we shifted from optimizing individual model runs to orchestrating full AI pipelines at runtime. That was a pivotal product evolution, and it completely changed how we approached the market.
We also:
- Secured our pre-seed funding through Unusual Ventures and onboarded into their Academy program.
- Filed foundational patents around dynamic inference and pipeline optimization.
- Launched from stealth and kicked off a private beta with design partners building real AI features in finance, productivity, and developer tools.
- And, perhaps most meaningfully, saw our first end-to-end runtime pipeline built live, from a real user prompt, with zero glue code. That was the moment we knew we’d built something fundamentally new.
Each milestone moved us closer to our mission: helping teams go from idea to adaptive AI feature, without orchestration headaches.
Pulse 2.0 (Amit): Are you able to discuss funding?
Skymel (Neetu): We’ve raised $500,000 in pre-seed funding from Unusual Ventures as part of the firm’s Unusual Academy and have an additional $25,000 in angel funding. These funds, combined with the capital we’re currently raising in our seed round, will be used to expand the engineering team so that we can accelerate NeuroSplit to market to serve the needs of our existing enterprise customers.
Pulse 2.0 (Amit): What total addressable market (TAM) size is the company pursuing?
The TAM for generative AI inference alone is projected to reach $168 billion by 2032 (Bloomberg), but we believe the opportunity is much bigger. We’re not just serving inference – we’re enabling the runtime layer that turns user prompts into working, adaptive pipelines.
We’re targeting the emerging category of AI execution infrastructure, which sits between model providers and AI-powered applications. Every company building with AI needs this layer. Whether it’s a startup embedding GPT APIs or an enterprise launching a product copilot, they all face the same orchestration, scaling, and integration challenges.
Skymel’s market includes every product team building intelligent features, every AI platform team struggling with pipeline complexity, and every developer trying to bridge the gap between prototyping and production.
Pulse 2.0 (Amit): What differentiates the company from its competition?
Skymel (Neetu): Skymel is focused on something most teams don’t have the time or tooling to build well: a runtime system that actually figures out how to connect models, tools, and data into a working AI feature, live, per user request.
A lot of the ecosystem is built around giving developers access to the ingredients – LLMs, APIs, data frameworks – but you still have to figure out how to wire those pieces together yourself. That’s where teams get stuck. We’re not trying to replace your models or host your infrastructure. We help you take everything you already use and turn it into something usable, dynamic, and production-ready.
Skymel (Sushant): Where we really stand apart is in how we handle orchestration. Other platforms let you set up static pipelines or choose the best model for a single step. Skymel goes further. The Orchestrator Agent looks at the full task, breaks it down, and creates a custom pipeline that considers how each step fits into the bigger picture. It’s not just about saving time or money at one point – it’s about getting the best outcome across the whole workflow.
Another big difference is how much control we give developers. You’re not throwing in a prompt and hoping for the best. You define what the feature is supposed to do, what it’s not allowed to do, and what trade-offs are acceptable. Skymel builds within those rules. You stay in control of the what – we handle the how.
And we don’t force teams to go all in on day one. You can start by using Skymel for one small part of your stack, like evaluation or post-processing, and expand from there as it makes sense. We’re not a framework you commit your whole product to. We’re a tool that helps you ship faster, with less overhead, and way more flexibility.
Pulse 2.0 (Amit): What are some of the company’s future goals?
Skymel (Neetu): Our long-term goal is to become the default runtime layer for intelligent applications. We want developers to focus on product intent, not orchestration, and have Skymel handle the rest.
In the short term, that means expanding OA’s capabilities: better observability, native support for agent memory and scratchpads, more out-of-the-box connectors to tools and APIs, and deeper developer UX improvements.
We’re also working on powering fully dynamic agent flows – where agents can reason, plan, and act with guardrails, all grounded in developer-defined feature intent. We believe Skymel can make agents safe, reliable, and production-ready for real business use cases – not just demos.
In parallel, we’re growing our ecosystem – partnering with toolmakers, hosting providers, and AI model platforms so developers can use Skymel wherever they already build.
At the core, we’re focused on a simple idea:
Let developers define what should happen. Let us figure out how.
