Machine Learning Engineer

Idara Samuel Osu

3+ years building and shipping ML systems in production — from model training and fine-tuning to inference optimisation, MLOps pipelines, and spatial forecasting. Built the first bilingual English-Ibibio AI platform from scratch for a language with 10M+ speakers and zero prior digital resources, and shipped a demand forecasting engine that predicts delivery zone activity one hour ahead using H3 spatial indexing and LightGBM on AWS SageMaker.

Nigeria · Remote or Hybrid · Open to full-time


What I build

LLM and speech systems

Fine-tuning, instruction tuning, NMT, TTS, ASR. HuggingFace, PyTorch, LangChain, LangGraph.

Inference engineering

TensorRT-LLM, vLLM, CTranslate2, KV caching, in-flight batching, FP8/INT8 quantisation. p99 latency reduction at production scale.

MLOps and cloud

AWS SageMaker Pipelines, Model Registry, Model Monitor, Endpoints. Modal Labs, Docker, GitHub Actions CI/CD, Terraform, S3, ECR.

Agentic pipelines

Multi-agent orchestration, tool calling, schema-constrained routing, compiler-composer architectures.


Experience

Head of Machine Learning

USEM AI · Nigeria
Jan 2024 — Jun 2026

Sole ML lead for the first bilingual English-Ibibio AI platform. Built every layer from scratch: data collection, corpus curation, model training, inference optimisation, and production deployment.

  • Collected and curated the first digital corpus for Ibibio: 80 hours of bilingual speech and 205,000 parallel sentence pairs. No prior dataset existed for this language.
  • Fine-tuned Whisper (ASR) and Orpheus TTS via LoRA on NVIDIA H100 GPUs. Achieved ASR word error rate of 15.33% and TTS naturalness score of 3.14 out of 5.
  • Adapted Meta's NLLB for English-Ibibio translation by transferring from Igbo, a related language. Gained 28 BLEU points over the zero-shot baseline.
  • Converted models to TensorRT-LLM and CTranslate2 with quantisation and KV caching. Cut p99 latency from 1,800ms to 320ms and GPU cost by 40%.
  • Architected a LangGraph pipeline with schema-constrained intent routing across 4 operations. Serving 250+ active users across translation, transcription, and audio generation APIs.

ML Engineer

Demz Analytics / Storri AI · Contract
Jul 2024 — Nov 2024

Built the LLM orchestration pipeline that converts raw news articles into structured content powering Storri AI's core product.

  • Chained approximately 30 GPT functions with output validation and JSON-repair logic, enabling unattended production operation without manual editorial correction.
  • Designed the MongoDB data layer storing and serving all AI-generated content for fast retrieval across the video-generation workflow.

Products

Self-directed builds shipped on personal initiative — designed and run on real production infrastructure.

EchoScript

Freelance · Automated Transcription Platform
Jul 2026

Automated transcription platform built on a fully SageMaker-native pipeline. Raw audio in, accurate timestamped transcript out, with no manual steps.

  • Whisper converted to TRT-LLM with paged KV cache. A threaded 20ms batcher collects mel tensors; a drain thread resolves futures by request ID.
  • 3-step SageMaker Pipeline: Download, Transcribe, GenerateDocs, with dependency-chained S3 I/O and containerised ScriptProcessors.
  • All infrastructure provisioned as code with Terraform: S3, ECR, IAM, CloudWatch, GitHub OIDC.
  • 6-stage GitHub Actions CI/CD: ruff, pytest, Terraform apply, Docker push to 3 ECR images, model upload, pipeline upsert.
  • Transcription time cut from 45 minutes to 8 minutes per audio-hour. Word error rate held at 6.8%. 850+ hours processed.

Projects

Lagos Demand Forecasting Engine

Spatial ML · AWS SageMaker
2026

Predicts food delivery order volume by neighbourhood zone one hour ahead. Translates that prediction into a rider pre-positioning recommendation — telling operations exactly how many riders to stage in each zone before demand spikes.

  • H3 spatial indexing pipeline converting raw coordinates into uniform 500m delivery zone cells
  • Feature engineering with lag features (1h, 2h, 3h, 24h, 7-day same hour), rolling averages, and time-of-day flags
  • LightGBM regression model with log-transform on the target — reduced prediction error by 68% over the baseline model
  • Full SageMaker pipeline: Processing Job, Training Job, Model Registry, real-time inference endpoint

Education

BSc Urban and Regional Planning

University of Uyo, Nigeria · CGPA 4.35 / 5.0
2017 — 2023

Spatial analysis, GIS-based mapping, infrastructure development, and Nigerian land law. Projects included site analysis and master plan preparation for urban growth across Nigerian cities — directly informing how I think about spatial data and zone-level modelling.