inioluwa.
🤗 Hugging Face

AI/ML engineer · MLOps · AI evaluation

Models thathold up outsidethe notebook.

I’m Inioluwa Atanda. I fine-tune models, build the harnesses that measure them honestly, and run them on infrastructure I can load-test and price.

$ training LoRA adapters on durable CPU workers; scoring Yoruba tone errors under quantization; red-teaming agents in Hausa, Igbo and Pidgin; autoscaling speech models on k3s with Helm; pricing every 1,000 inference requests

See the products 🤗 Run a model in your browser
Case studies
17
Open research
4
HF releases
6
Kaggle
Contributor

Fine-tuning✦LoRA / PEFT✦Evaluation harnesses✦Multilingual safety✦Speech recognition✦Quantization✦Model serving✦Kubernetes✦Observability✦Cost per 1k✦Agents with guardrails✦Reproducibility✦

How I work

The whole lifecycle,
not just the training run.

Scroll the five stages. The field behind this page turns into each one: data clusters, a network, AdapterLab’s real confusion matrix, a serving cluster and a latency surface.

  1. 01

    Curate

    Data

    Audit corpora, remove leakage, fingerprint versions and split by group so a test set stays a test set.

    • Dataset audits
    • Group-aware splits
    • Fingerprints
    • PII flags
    ProofWAXAL Yoruba audit1.56 h held-out test audio
  2. 02

    Adapt

    Train

    Fine-tune with LoRA or full updates on durable, cancellable jobs; choose checkpoints on validation, never on test.

    • PyTorch
    • PEFT / LoRA
    • Whisper
    • Durable workers
    ProofAdapterLab0.204% trainable parameters (LoRA)
  3. 03

    Measure

    Evaluate

    Build harnesses that report failures next to wins: per-class scores, tone errors, multilingual safety and paired uncertainty.

    • Harnesses
    • Judge calibration
    • Baselines
    • Error analysis
    ProofSafety across languages16,000 planned test cases
  4. 04

    Ship

    Serve

    Quantize, export and deploy behind authenticated gateways on Kubernetes, with limits, probes and autoscaling.

    • ONNX
    • int8 / q4
    • vLLM
    • k3s + Helm
    • Terraform
    ProofAfrican-language serving$0.139 per 1k requests, measured
  5. 05

    Operate

    Observe

    Load-test, break pods on purpose, watch the metrics and price every thousand requests before calling it done.

    • Prometheus
    • Grafana
    • k6
    • Cost per 1k
    ProofCPU or cloud?3,400 parallel eval items

Selected products

Things I built,
with the numbers attached.

Eunini/adapterlabMLOpsAdapterLabFine-tuning as a repeatable product.A LoRA fine-tuning workbench with dataset validation, durable training jobs, held-out evaluation, export and inference.0.204%trainable parameters (LoRA)PyTorchPEFT / LoRAFastAPIONNXReactOpen case study
Eunini/maooAgentsMAOOAutonomy needs a control loop.A multi-agent orchestrator that perceives, plans, executes and recovers with explicit limits.24evaluation scenariosPythonPydanticSQLiteDockerOpen case study
Eunini/afri-ai-servingMLOpsAfrican-language servingGive the model somewhere dependable to run.A measured serving platform with infrastructure as code, authenticated gateways, autoscaling and failure tests.$0.139per 1k requests, measuredTerraformk3sFastAPIPrometheusOpen case study
Eunini/yoruba-asr-tone-compressionResearchYoruba tone & compressionWhen a model gets smaller, what disappears?Studying the relationship between speech-model compression, word recognition, Yoruba tone marks and serving cost.1.56 hheld-out test audioPyTorchWhisperCTranslate2whisper.cppOpen case study
Eunini/nigerian-agent-safetyResearchSafety across languagesA safety rule should survive translation.A controlled study of refusals, harmful completions and simulated tool attempts in Nigerian languages.16,000planned test casesAgent evaluationHarmBenchXSTestPythonOpen case study
Eunini/huggingface-portfolioAgentsVoice-agent workflow labsMake an agent’s decisions replayable.Two public, inspectable simulators for voice-agent workflows: an AI SDR and a hotel receptionist, backed by a scenario dataset.34public regression scenariosTypeScriptState machinesTool tracesHugging FaceOpen case study

Open research

Four questions about AI
for African languages.

Each study is public, in progress and labelled that way. Finished phases are filled in; nothing unfinished is reported as a result.

R1When a speech model gets smaller, do the tone marks go first?Yoruba tone & compression · Research in progressCorpus auditFine-tune ×3Compression scoringServing cost
R2What does one Yoruba, Hausa or Igbo answer really cost?CPU or cloud? · Research in progress3,400-item setGPU runCPU + API runsCost model
R3Can a small African-language model service hold up under load?African-language serving · Measured prototypeTerraform + k3sLoad & chaosFine-tuned imageSign-off
R4Do an agent’s safety refusals survive translation?Safety across languages · Protocol & implementationProtocolInert toolsNative reviewMain runs

Eval lab

Evaluation is the product.
Play with the real results.

Every chart below is drawn from a recorded run, failures included. Switch runs, compare languages, find where the model breaks.

true ↓ / pred →accobillcancdelilogitech
account5
billing5
cancellation5
delivery32
login41
technical5
Billing is read as cancellation and cancellation as delivery. Both score zero F1, and the model card says so.

0.520macro F1

0.567accuracy

30held-out examples

  • account1.00
  • billing0.00
  • cancellation0.00
  • delivery0.46
  • login0.89
  • technical0.77
How AdapterLab produced this

Everything

The full index.

Filter by area or search the stack. Press / anywhere to open the command palette.

17 projects

  1. 01AdapterLabA LoRA fine-tuning workbench with dataset validation, durable training jobs, held-out evaluation, export and inference.MLOpsPyTorch · PEFT / LoRA · FastAPIPublic source
  2. 02MAOOA multi-agent orchestrator that perceives, plans, executes and recovers with explicit limits.AgentsPython · Pydantic · SQLitePublic source
  3. 03AgroPredictA crop-yield prediction application connecting a random forest to farmer and administrator workflows.Applied MLscikit-learn · Flask · SQLAlchemyPublic source
  4. 04ReproPackA reproducibility tool for packaging dependencies, setup scripts and project metadata into a shareable archive.MLOpsFastAPI · Pydantic · PythonPublic source
  5. 05Yoruba tone & compressionStudying the relationship between speech-model compression, word recognition, Yoruba tone marks and serving cost.ResearchPyTorch · Whisper · CTranslate2Public source
  6. 06CPU or cloud?Benchmarking Nigerian-language question answering across small CPU models, a larger GPU model and a free-tier API.Researchllama.cpp · vLLM · IrokoBenchPublic source
  7. 07African-language servingA measured serving platform with infrastructure as code, authenticated gateways, autoscaling and failure tests.MLOpsTerraform · k3s · FastAPIPublic source
  8. 08Safety across languagesA controlled study of refusals, harmful completions and simulated tool attempts in Nigerian languages.ResearchAgent evaluation · HarmBench · XSTestPublic source
  9. 09Voice-agent workflow labsTwo public, inspectable simulators for voice-agent workflows: an AI SDR and a hotel receptionist, backed by a scenario dataset.AgentsTypeScript · State machines · Tool tracesPublic source
  10. 10GridSenseOffline-first solar, battery and grid monitoring with durable replay, MQTT ingestion, forecasting and explainable anomaly alerts.Applied MLPython · FastAPI · MQTTPublic source
  11. 11Mule-ring detectorA Rust temporal-graph detection engine connected to a Java investigation desk for synthetic financial transactions.Applied MLRust · Gradient-boosted trees · JavaPublic source
  12. 12Helios AIA multimodal business-agent application with voice, video and text interfaces over a FastAPI backend.AgentsFastAPI · Next.js · WebSocketsPublic source
  13. 13CUDA ForgeA portable agent skill for setting up and checking NVIDIA GPU inference infrastructure on Linux.MLOpsLinux · CUDA · DockerPublic source
  14. 14Voice reservationsA Nigeria-focused hotel voice-AI booking platform built around realtime model interaction.AgentsVoice AI · Realtime APIs · Backend systemsPrivate source
  15. 15NoiseGuardAn edge-audio monitoring application linking a C capture agent with alerts, history and reporting.Developer toolsC · PortAudio · SQLitePublic source
  16. 16FastAPI cloud deploymentA deployment template connecting a Python API to container builds, infrastructure and AWS operations.InfrastructureDocker · Ansible · TerraformPublic source
  17. 17AWS network foundationsTerraform networking work covering segmented infrastructure and the services deployed inside it.InfrastructureTerraform · AWS VPC · LinuxPublic source

Toolbox

Skills, by layer
of the stack.

Find me

Models, notebooks
and source.

Hugging Face@Ineeoluwa

Models, datasets and Spaces you can run in the browser, each with a card that states its limits.

1
model
2
datasets
3
Spaces
Kaggle@inioluwaeunice

GPU notebooks for benchmark runs, including the full 3,400-item Nigerian-language evaluation.

Contributor
tier
2+ yrs
on Kaggle
9
badges
  • Python CoderBadge
  • Dataset CreatorBadge
  • API Notebook CreatorBadge
  • API Dataset CreatorBadge
  • Colab CoderBadge
GitHub@Eunini

Source for the products, research harnesses and infrastructure on this page.

89
public repos
4
research repos
17
case studies

Profile figures are a snapshot from 28 October 2026.