UUAMNI Incorporated · Delaware Public Benefit Corporation
UUAMNI Energy · UUAMNI AI Nigeria Ltd · CAC Nigeria Registered
Tier III Lagos Datacenter
US Export Approval Obtained · June 2026
"Ulunma, ụgbọ ala m na-eweta Ìhè!"
Building the labor, training data, evaluation, and compute infrastructure that AI structurally requires to evolve across the globe.
AI cannot hold all intelligence if it cannot think in every language.
Starting in Africa. Igbo, Yoruba, Hausa, Pidgin, and beyond.
Model serving, fine-tuning, and inference on infrastructure in Lagos, with per-tenant isolation, audit export, and a signed statement of where each workload ran that your own engineers can verify. Running today in the AWS Lagos Local Zone; owned hardware on a Tier III floor in Lagos is next.
See the runtime →For labs and researchersUche produces Igbo-origin preference pairs with native-speaker judgment, reliability statistics, and provenance on every row. IlùBench, our published benchmark, is public with an open-source runner and a live leaderboard.
See the evidence →01 / The Dependency
Every AI system on earth runs on human feedback. The question is where that human input comes from, and who controls the supply.
Safety refusal collapse from ~90% in English to 35–55% in Yoruba, Hausa, Igbo, and Igala (worst case shown). Frontier-model alignment does not transfer to low-resource African languages.
Contracted power for sovereign AI infrastructure at a Tier III Lagos datacenter. GPU capacity built to be owned on the continent.
Published Evidence · Since July 2026
Ask a frontier model to explain an Igbo proverb in English, then in Igbo. It reasons from a different cultural frame each time. IlùBench is the first benchmark measuring cultural register switching in an African language: a published protocol, a seed probe set with attestation tracked per probe, a scoring rubric anchored in native-speaker judgment, and evidence across frontier and open-weight models on a public leaderboard. One open-weight model answered Igbo prompts in fluent Yoruba, a failure class nobody had documented. Run it on any model yourself with our open-source runner.
Benchmark on HuggingFace →|Open-Source Runner →|Leaderboard →
02 / The Stack
One factory, three surfaces, one site. If any layer is separated, someone else controls whether the teaching continues. Each layer finances and improves the next. Data is finite. A data factory compounds.
The cultural-IP layer that produces preference data no remote workforce can produce. UUAMNI recruits native speakers with the cultural depth to score linguistic nuance, work that requires literary fluency the job cannot teach. The first cohort is Igbo; the method is the same for every language after it. Cultural judgment has no credential to verify, so we built the verification system that makes it auditable: every dataset ships with the reliability statistics to prove it. Margin structure flows from data-product pricing, not from labor arbitrage. The labor universe incumbent data vendors cannot recruit is the one we sit on top of.
In-country, supervised, auditable. OpenAI-compatible inference and model serving, fine-tuning, per-tenant isolation, metering and billing, and audit export, with a signed statement of where each workload ran that a bank's own engineers verify in their browser. Fraud scoring and AML monitoring with reason codes, evaluable on synthetic data in about an hour. Running today in the AWS Lagos Local Zone on the deployment automation that will run on owned A100 hardware at a Tier III datacenter in Lagos, co-located with the annotation workforce. Reserved and wholesale tiers for labs, banks, and regulated institutions.
The first open Igbo-origin DPO dataset. v1 target: 20,000 strong-preference pairs across eight tracks, from translation and cultural QA to safety and refusal. Native-origin Igbo, not translated English. Every release ships with benchmark evidence, measured against IlùBench, our published evaluation layer. CC-BY-4.0 public sample at launch, commercial tier for frontier labs. The data layer that closes the safety floor and opens the cross-lingual ceiling at once.
Fair wages flowing directly into local Nigerian economies. Skills development creating long-term careers. Long-term roadmap: solar-powered, owned-and-operated GPU capacity on the African continent. Excess capacity distributed to communities, hospitals, and schools as the build matures.
03 / Work With Us
One stack, four ways in. Tell us which one you are.
Native-origin African-language preference data with an evidence package, plus sovereign compute to train on. Talk to us →
Nkwurite: fine-tuned fraud, credit, and document models served on infrastructure in Lagos, with per-tenant isolation and a signed statement of where each workload ran. Data stays in Nigeria, with an NDPC-registered operator. Run the evaluation → · Talk to us →
Fair-wage, expert-level language work: register, proverb, and cultural reasoning. Careers, with real employment terms. Join the workforce →
Native preference data, African-owned compute, one integrated stack. Get in touch →
05 / Why Now
No one has done this work for any African language yet.
Meta paid $14.3B for its stake in Scale AI, for access to human feedback. The constraint is genuine human intelligence at scale. The companies that secure this input win. The ones that don't, stall.
No intelligence is formed in isolation. Every human civilization evolved through contact with difference. AI trained without African languages, reasoning patterns, and perspectives is a mirror admiring its own reflection. A universal intelligence needs contact with difference. The structural gap is growing, not shrinking.
RLAIF, AI evaluating AI, produces model collapse. Systems trained on their own output degrade. The need for genuine human signal is permanent.
Nigeria's central bank now requires every bank, fintech, and payment processor to store and manage all payment transaction data inside Nigeria by January 1, 2027, with sanctions for non-compliance, and most of these institutions run on foreign cloud today. As of October 2026, no other supplier in Nigeria publishes a verifiable record of where each AI workload ran. Compute follows the data. NITDA's 2026 cloud guideline goes further for banks: their financial data must be “primarily hosted, processed and stored” in Nigeria. UUAMNI holds the workforce, the data, and in-country compute as one integrated stack.
Imo State alone has trained 50,000 young people in digital skills through SkillUp Imo, with the literacy and English-fluency foundation that supports advanced annotation training. UUAMNI builds on top of that base with the additional cultural-IP layer, scholars, native-fluency annotators, and PhD-level linguistic expertise, that frontier-lab preference data actually requires.
07 / The Founders
Two engines: an operating company and a public-intellectual research engine, compounding in parallel.
Co-Founder & Chief Executive Officer
Nigerian-American, based in New York City, operating between NYC and Lagos. 13+ years building enterprise infrastructure for financial services: four years at Microsoft, then nine-plus years in capital-markets fintech supporting the platforms effectively every major US investment bank relies on to run IPO order books. Leads UUAMNI's technical, commercial, capital, and Nigerian-operations stack.
Co-Founder & Chief Strategy Officer
PhD Princeton Theological Seminary, Spring 2026. Architect of UUAMNI's structural-necessity thesis: if AI is ever to be a universal intelligence, it must have access to ways of reasoning beyond what it is currently trained on. That starts with Africa. Her research program, ontology, model collapse, and the limits of alignment and synthetic training, births the ideas at the root of UUAMNI's products and pushes the conversations that put the company at the frontier of today's AI discourse. Her position paper, The Closure Problem in Alignment, is in submission.