Independent product studio · Georgia
OguzHub is a small studio working from the Caspian corridor. We take products the whole way — model layer, backend, infrastructure, and the app in your hand.
Every claim on this page is checkable. That’s deliberate.
Products
We build for ourselves first. Each of these went from an empty repo to something running — one judged against 10,000 entries, the other against an adversary trying to break it.
A behavioural-change engine disguised as an alarm clock.
An algorithm walks your wake time earlier on a push–push–hold–recover cycle. You don't get to set the time; that's the point. When you dismiss it, a speech written for the identity you chose plays back in your own language.
AWS 10,000 AIdeas — Top 50 finalist, from 10,000+ entries
Documents in. A cited compliance verdict out.
A product-compliance reviewer. The model reads and extracts; a deterministic rules engine makes every decision. Each flag names the rule it broke — so a human checks the work instead of trusting it. Published as an Apify Actor.
7/7 attacks defended in our own adversarial red-team battery
Agents
Not demos. These are the pieces the two products above are actually made of — and the engineering that keeps each one dependable when it's the only thing awake at 4am.
Reads every user's history each night, decides push / hold / recover, computes tomorrow's alarm, and folds in sunrise and weekend offsets.
Deterministic end to end: the same history always yields the same alarm, and no part of tomorrow's time is a model's opinion. A user can be told exactly why it moved.
Bedrock writes a speech for the user's chosen identity in their own language; Polly voices it and it lands as audio before the alarm rings.
Curated content for days 1–30 skips the model entirely. Every generation checks S3 first, so a re-run costs nothing. Throttling backs off adaptively and re-raises, so orphaned jobs replay instead of vanishing.
A second content line — evening passages with a chapter arc, rated per chapter, delivered as audio.
Runs off the same fan-out as the morning line: one orchestrator writes the payloads, a pool of workers renders them.
Reclassifies every user across Standard, Struggling, Mastery, Trust and Graduated on their own success and engagement rates.
Thresholds are explicit and deterministic, not a model's opinion — a user can be told exactly why their state changed.
Maps a free-form title someone types into one of eight archetypes, and screens custom identities before they're accepted.
Structural checks plus a denylist run before anything reaches the model, so user input never becomes a prompt.
Reads whatever a supplier sends — PDF, photograph, pasted text — and pulls structured facts out of it.
A Bedrock Guardrail pins it to extraction. It has no verdict field to fill in, so it cannot express an opinion even when a document asks it to.
Checks every extracted claim actually appears in the source, and strips verdict keys that a document tried to smuggle in.
This is the anti-hallucination gate. Anything the model asserts that isn't in the document doesn't survive it.
Makes the actual call — approve, needs review, reject, quarantine.
A deterministic rules engine, not a model. Every flag names the rule it broke, so a reviewer checks the reasoning instead of trusting it.
Attacks our own agent with seven injection techniques: hidden HTML, comments, unicode tags, base64, smuggled verdicts, visible instructions, prompt extraction.
7/7 defended. We run it against ourselves because a claim about security that nobody tried to break isn't a claim.
When documents are missing, drafts the email that asks for them.
Human-in-the-loop by design: it drafts, it never sends. A person approves every message that leaves.
Approach
A studio is its judgement. These are ours, and they show up in the code.
Language models are extraordinary readers and unreliable judges. Ours pull structure out of messy input, then hand off — deterministic code makes the call. That's why our agents can cite the rule behind every verdict instead of asking you to trust a black box.
Model layer, backend, infrastructure, and the app in someone's hand. A studio that only does one slice spends its life arguing across a boundary. We'd rather own the boundary.
No roadmap dressed up as a product, no pilot called a partnership, no number without a source. It costs us some swagger. It buys something worth more — when we say a thing works, you can go check.
Capabilities
Not a menu of everything that exists — the things we've shipped something real with.
About
The Oguz were the Turkic confederation whose migrations seeded Anatolia, Persia, and Central Asia — a corridor of movement and trade running from the Caspian outward. Naming a software studio after them is a claim about where the next wave gets built: the same corridor, a thousand years later.
There is an enormous amount of engineering talent along that belt and almost no infrastructure pointing it at global demand. We’re building products from inside it rather than writing papers about it.
To be plain about what we are: OguzHub is an independent studio. We’re not an accelerator. We don’t take applications, we don’t invest, and we don’t run programmes. We build our own products and ship them.
Contact
We’re most useful on problems where an AI agent has to be right, not just fluent — and someone has to be able to check it afterwards. If that’s the shape of your problem, write to us.
