// AI intelligence optimization
Nandromeda is building the optimization layer that sits between AI applications and the models they call — deciding, per request, the cheapest path to a result that's actually good enough. Xentient is the first product on that thesis: a router that quantizes, distills, and re-serves the models you already run, then checks its own work before it ever reaches you.
Early access for teams spending real money on inference.
Same answers. Same quality. Just a smaller bill.
That's the whole idea, in plain terms: Xentient doesn't make your AI dumber to save you money. It just stops sending simple tasks to expensive models that never needed to see them.
Illustrative breakdown of where inference spend typically leaks — not a guarantee for any specific workload.
Reduced numerical precision where the workload can absorb it, tested against your own eval set.
For narrow, repeated tasks, a distilled model can match production output at a fraction of the cost.
Serving infra tuned for the traffic pattern it actually sees.
A rough, directional estimate — not a quote.
One illustrative request, run two ways.
Indexed to GPT-4o direct = 100. Illustrative, not a live benchmark yet.
*cheapest-always fails hard, ambiguous, or high-stakes tasks — routing exists so you don't choose between cost and correctness.
Xentient is Nandromeda's flagship — a task router that delegates each LLM call to the most economical model that can still handle it well, then checks its own answer before it ever reaches you. It's the entry point into a bigger thesis: the cost that matters isn't per token — it's per successful task.
Tell us about your setup and we'll walk you through how routing would work for your traffic.
We'll reply personally — no auto-generated sales sequence.
Not yet — we're pre-launch. Join the waitlist above and we'll reach out when there's a pilot slot open.
Depends on your traffic mix. Figures on this page are illustrative, not a guarantee.
The plan is broad multi-provider support — GPT, Claude, Gemini, Llama, Mistral, and others.
No formal certifications yet — too early for that. A real data-handling policy will be published before general availability. See our page.
No — the pitch works whether or not you're technical: same output, lower bill. Engineers can go deeper on our page.
Xentient sits in front of your model calls, reads what each task actually needs, and routes it to the most economical model that can still do the job — then checks whether the answer actually cleared the bar before it reaches you. So your expensive model only fires when the task earns it, and cheap only wins when cheap actually worked.
A simplified illustration — the real classifier weighs risk, ambiguity, and stakes, not just sentence length.
Xentient is Nandromeda's premier flagship product — a task router that sits between your application and the LLMs you already use. Every request that would normally go straight to one fixed model passes through Xentient first.
It looks at what the task actually needs — complexity, ambiguity, risk — and forwards it to the model tier suited to that request, instead of defaulting everything to your most capable (and most expensive) model. Then it checks the result. A cheaper model that fails and needs a human to fix it isn't actually cheaper — so the number we're optimizing for is cost per successful task, not cost per token.
The exact scoring, evaluation, and routing logic is where the real work happens, and it's the part we're keeping close for now. What we can say: it's built to get out of your way — one integration point, decisions made per request, nothing for you to manage day to day.
Every call gets classified on complexity, ambiguity, and how much a mistake would cost — then delegated to the tier that matches it.
Complexity, ambiguity, risk, and the quality bar the answer has to clear — before any model is chosen.
Not the cheapest model, period — the cheapest one with a real shot at getting this specific task right.
An answer that looks fine and an answer that's actually right aren't always the same thing. This is the step routing-only tools skip.
You get the cost savings when they're real, and the accuracy floor when they're not — instead of choosing one or the other upfront.
Honest status: this evaluate-and-escalate loop is the design we're building toward, not a live benchmark yet. It's the difference between "we picked a smaller model" and a cost claim you can actually trust.
One integration point in front of your existing model calls. Xentient doesn't replace your models — it decides, per request, which of the ones you already use should answer. You keep your accuracy floor; it just stops your simplest traffic from hitting your most expensive endpoint.
Xentient is being built alongside its first pilot teams.
No spam — just a note when there's a pilot slot open.
Nano changes. Andromeda outcomes.
MISSION
We exist to strip the waste out of running AI in production — starting with inference. Every task routed to a model bigger than it needs is a small tax on innovation. Our mission is to make sure that tax stops deciding what a team can afford to build.
VISION
We think about this the way engineers think about compilers: a layer beneath the code that optimizes automatically. We want to be that layer for intelligence — starting at model-routing with Xentient, reaching over time toward the logic and chip layer beneath it. Not a bigger model. A smarter layer underneath all of them.
ORIGIN
The name stacks three ideas: Nano, the smallest scale of change. NAND, the logic gate that's functionally complete on its own. And Andromeda, our nearest large galactic neighbor, roughly a trillion stars. The smallest, most unglamorous optimizations compound into differences too large to ignore.
ARTICLES · COMPANY THESIS
Most teams pick one model per feature and stop thinking about it — a reasonable default, and a quietly expensive one. The model isn't the product decision — the routing is. Xentient exists to make that decision as deliberate as the decision to build the feature at all.
A NAND gate is almost nothing by itself — a handful of transistors — yet every digital system reduces to compositions of gates like it. One request routed correctly saves a fraction of a cent. Multiply that across every request a growing product makes, for years, and the fraction becomes the difference between an infrastructure bill being background noise or an existential cost.
ROADMAP
Directional, not a fixed schedule.
Now
Xentient routing engine + pilot waitlist
Building the core router and onboarding early pilot teams.
Next
Context & caching layer, pricing finalized
Reduce repeated-context cost, lock in gainshare-style pricing.
Later
Developer-surface expansion
A lightweight editor extension bringing routed completions into existing workflows.
Moonshot
Independent enterprise LLM deployments
Private, efficiency-optimized deployments for companies wanting independence from big AI vendors.
CAREERS
No formal open roles yet — Nandromeda is early and small. But if cutting the waste out of running AI in production is a mission you want to help build, we want to hear from you.
Especially useful right now: ML/infra engineers, and anyone who's shipped LLM products in production and felt the cost pain firsthand.
Tell us what you'd want to work on — no formal application needed at this stage.
PRESS KIT
Logo files and extended brand assets available on request — email hello@nandromeda.com.
FOR INVESTORS
We're an Istanbul-based team building the intelligence optimization layer for AI workloads — the decision layer that sits between applications and the models they call. Xentient, our inference-cost router, is the first product on that thesis.
Full-screen, ten slides, two minutes. The written version is below either way.
THE PROBLEM
Teams default nearly all traffic to their single most capable model, regardless of what each task actually needs. That compounds silently until the AI infrastructure line item rivals headcount — and it's getting worse, not better, as agentic workflows call models more often and in longer chains. Gartner projects inference cost per agentic workflow could increase more than fivefold through 2028, even as per-token prices keep falling (Gartner, Aug 2026). AI gets cheaper per token and more expensive per business outcome at the same time — that paradox is the market.
THE SOLUTION
Xentient sits in front of existing LLM calls and routes each one to the most economical model that can still handle it, then checks whether the answer actually cleared the bar before accepting it. No model switching for the customer, no retraining, one integration point. The metric we're building toward isn't tokens saved — it's cost per successful task: a cheap model that fails and needs a human fix isn't actually cheap.
MARKET
Source: Precedence Research, AI Inference-as-a-Service Market (2026). Enterprise GenAI spend separately tripled from $11.5B to $37B year-over-year (Menlo Ventures, via Hostinger LLM Statistics 2026), and 73% of enterprises now spend over $50K/year on LLM APIs (Kong Inc. survey, via Calliber). Xentient addresses the optimization layer sitting on top of this spend, not the whole market — even a small share of it is a large opportunity.
WHY THIS DOESN'T STAY A ROUTER
Routing alone is becoming commoditized — RouteLLM is open source and already reports up to 85% cost reduction while retaining ~95% of GPT-4-tier performance on its benchmarks (RouteLLM, lm-sys), and OpenRouter and LiteLLM already offer multi-model gateways. A fair investor question is "why can't they just add this?" Our answer is that what compounds isn't the router — it's what it learns sitting in front of real traffic:
Which model actually succeeds on this customer's workload, not benchmarks in general — that's specific to each deployment and gets better the longer we run.
Judging whether an answer was actually good enough is a harder, separate problem from picking a model — and it's the part that makes any savings number trustworthy.
Capability, price, and reliability across providers change constantly. Staying current on that is its own ongoing job.
Data residency, approved providers, and stakes-based rules (e.g. legal tasks always escalate) become part of the routing decision, not an afterthought.
BUSINESS MODEL — directional, not finalized
The plan is a platform fee sized to the customer, plus — for larger deployments — a share of verified savings: measured against a baseline of what the customer was spending before Xentient, so the pitch is arithmetic, not a promise. If we save a customer real money, that's provable and the pricing conversation follows from the number, not the other way around. Nothing here is a signed term sheet — it's the shape we're building toward.
TRACTION & STAGE
Domain, brand, and product direction are set. Heads-down on the routing engine and early access rollout. Not currently running a formal raise, but always glad to talk to investors thinking about this space early.
THE ASK
$900K on a SAFE — $7.5M cap
Introductions to early pilot customers, and investors who think about deep-tech / AI-infra efficiency plays are all useful right now — capital included, but not the only thing we're after at this stage.
Pre-seed SAFE, sized to where we actually are — pre-product, pre-revenue. Not a priced round, and not financial or legal advice; terms subject to counsel.
GET IN TOUCH
investors@nandromeda.comWe read every note personally. Ask and we'll send the full pitch deck PDF directly — it's not hosted publicly at this stage.
// for engineers
Route every prompt to the model that actually fits it — on cost, complexity, and risk. Below is the current design, not a live API yet.
A simplified simulation — adjust the constraints and see how the decision shifts.
Illustrative simulation, not connected to a live model or real cost data.
Tracks per-provider token pricing and shifts load toward cheaper models for repetitive, low-stakes tasks.
Flags PII, high-liability, or ambiguous prompts and escalates them automatically.
If a provider is rate-limited or down, requests fail over to an alternative model.
Planned: periodic sampling of routed outputs against your own eval set.
This is the interface we're designing toward — not a published package yet.
Before today
# hardcoded to one model, regardless of task from anthropic import Anthropic client = Anthropic() response = client.messages.create(model="claude-opus-4", messages=[...])
After planned
# same call shape, routed automatically from xentient import SmartClient client = SmartClient(api_key="xt_...") response = client.generate(prompt="...", cost_sensitivity="high")
Honest status: this is the target interface, so migration is a client swap, not a rewrite. It isn't installable yet — join the waitlist on the Xentient page for early access when it is.
Intent is zero unnecessary retention of prompt content. A formal policy lands before general availability.
We do not hold SOC 2, HIPAA, or similar certifications yet — pre-launch. We won't claim them until real.
No published benchmark yet. We'll share measured numbers once we have pilot traffic.
Designed in from the start — the whole premise is routing across providers.
Last updated: September 26, 2026
Heads up: Nandromeda is a pre-launch startup. This is a plain-language starting draft, not final legal advice — it should be reviewed by a lawyer familiar with Turkish (KVKK) and, if we take on customers elsewhere, EU (GDPR) data law.
We collect an email address only if you submit it via a waitlist/demo form, plus a single local preference flag for the cookie banner that never leaves your browser. No analytics, ad pixels, or trackers run on this site today. Waitlist emails are used only to contact you about early access — never sold or shared.
Ask us to delete your email at any time via hello@nandromeda.com. GDPR/KVKK rights apply where relevant.
Figures shown on this site (cost savings, latency, routing tiers) are illustrative unless stated otherwise, and are not a guarantee of results. This site and any pre-release product are provided "as is" during early development. Governed by the laws of the Republic of Türkiye, pending formal incorporation and legal review.
We use exactly one piece of local storage: a flag remembering your cookie-banner choice. Stored only in your browser, not used for tracking or advertising, shared with nobody. If we add analytics later, this policy and the banner will be updated first.