👋  I'm Idara — ML Engineer from Nigeria, building AI for languages the world forgot to include
🔭  Currently scaling USEM AI and teaching Etimbuk (my AI agent) her next language — Ibibio,Annang,Oro
🌍  On a mission to make every indigenous African language speak through AI
⚡  Fun fact: I love the Ancients — Isaac Newton and Nikola Tesla are my people
📫  idaraosusamuel@gmail.com  ·  @Osuaidy

Training          PyTorch · HuggingFace Transformers · NLLB · Whisper · OrpheusTTS · Molmo · H100
Inference         TensorRT-LLM · CTranslate2 · vLLM · SGLang · ONNX · fp16 · int4 AWQ
Optimisation      KV Caching · In-flight Batching · PagedAttention · Dynamic Shape Profiles
MLOps             Vertex AI Pipelines · SageMaker Pipelines · KFP · DVC · MLflow · Evidently AI
Agents            LangChain · LangGraph · Multi-agent Orchestration · Tool Calling
Serving           FastAPI · SSE Streaming · SageMaker Endpoints · Serverless · Docker
Monitoring        BigQuery · CloudWatch · Prometheus · Grafana · Drift Detection · Auto-retrain
CI/CD & IaC       GitHub Actions · Terraform · EventBridge · SQS · ECR
Data              MongoDB · PostgreSQL · Supabase · Proprietary Data Pipelines

USEM AI — Founder & ML Engineer · 2024–2026

Built a production multilingual AI platform for Ibibio, Annang, Oro, and Ekid from nothing — proprietary datasets, model training on H100s, full inference stack, and a live agentic layer. Converted TTS/STT to TensorRT-LLM and NMT + Whisper to CTranslate2. Applied KV caching, in-flight batching, and quantization for real-time throughput. Shipped Etimbuk — a LangGraph agent routing across translation, TTS, STT, and creative writing with proactive WhatsApp/Telegram notifications. 100+ active users.

Auto Insurance Damage Inspection — MLOps · GCP · Vertex AI

Vision pipeline for vehicle damage classification and severity scoring. Fine-tuned Molmo-2B as the backbone; full pipeline containerised on Vertex AI with DVC dataset versioning and MLflow experiment tracking. TensorRT fp16 conversion with per-head output diff assertion. KFP conditional deployment gate blocks promotion if accuracy < 0.85 or macro F1 < 0.80. Evidently AI drift monitoring on BigQuery inference logs triggers automated retraining when drift share exceeds 20%.

Production TTS — OrpheusTTS + TensorRT — Inference Engineering · AWS · SageMaker

OrpheusTTS → ONNX → TensorRT: encoder · decoder · HiFi-GAN vocoder, all fp16 with dynamic shape profiles. Sentence-chunked streaming via FastAPI SSE. SageMaker g5.xlarge deployment with 10% canary → smoke test → full traffic promotion in GitHub Actions. Async DNSMOS P.835 scoring on sampled outputs; RTF, TTFC, and GPU utilisation tracked via CloudWatch.

YouTube Summariser — Whisper + Mistral on One GPU — Inference Engineering · AWS

Whisper large-v3 and Mistral 7B, both TensorRT-LLM, co-deployed on a single A10G (24GB). Mistral quantized with int4 AWQ calibrated on CNN/DailyMail. PagedAttention + continuous batching + chunked context on the LLM. Videos over 10 minutes handled via SQS async queue. Prometheus tracks Whisper RTF, LLM TTFT, tokens/sec, and queue depth.

Churn Prediction — Postgres to SageMaker — MLOps · AWS · Supabase

Full SageMaker Pipeline DAG: SQL feature engineering in a Processing Job → GPU-accelerated XGBoost with SHAP → Model Registry approval gate (AUC ≥ 0.82, F1 ≥ 0.75) → endpoint deployment with live Supabase feature fetch at inference time. EventBridge cron → Lambda → Evidently AI drift check on live database → automated pipeline retrigger on drift share ≥ 30%.