Emmanuel Mayowa Samuel
AI/ML Engineer
I build AI that survives production. Agentic systems on AWS Bedrock and Google Vertex AI with MCP tool orchestration and RAG, classical ML for anomaly detection and forecasting, and the full stack around both - from PostgreSQL schema to deployment.
Emmanuel Mayowa Samuel
AI/ML Engineer
About
Emmanuel
AI/ML Engineer focused on LLM infrastructure, autonomous agents, and retrieval-augmented generation systems.
I build agentic LLM systems on AWS Bedrock and Google Vertex AI, where reliability, retrieval quality and cost all matter at the same time — and AIOps, where the models are looking at telemetry rather than language.
I build AI systems that run in production, not prototypes. Much of my work is agentic: LLM systems that use tools, hold state, and have to keep working when a provider returns a 429 at 3am - multi-tenant platforms on AWS Bedrock and Google Vertex AI, with guardrails, retries and per-call cost instrumentation. Plenty of it is not generative at all. I have built AIOps over live telemetry, with anomaly detection, forecasting, change-point detection and root-cause ranking, and ensemble models combining Poisson regression, XGBoost and Elo with Monte Carlo simulation. I own the stack around both, from PostgreSQL schema through FastAPI or NestJS to the Next.js surface and deployment.
Recognition
Google Solution Challenge — Minna Champions
Led the campus Google Developer Student Clubs chapter and the team that placed Minna Champions in the 2023 Google Solution Challenge, building an AI-powered waste management platform aligned to SDG 12. Also finished fourth as team lead at PAYAZA Hackathon 3.0, on an AI solution for MSME financial inclusion.
Architected LangGraph agents on AWS Bedrock AgentCore across 7 product lines, loading tool catalogues at runtime over MCP from vendored OpenAPI contracts.
Diagnosed a production tool-routing failure and re-architected a supervisor plus five sub-agent system into a flat single-agent design over a unified ~23-tool catalogue.
Built per-call AI cost metering with OpenTelemetry counters and a PostgreSQL audit log, pricing prompt, output, thinking and image tokens separately against a per-model rate table.
Cut OPEX report preparation from one week to two days at a power plant, and cleared a standing ticket backlog across a 430-bed estate serving 500+ staff.
Professional
Experience
Trained as an electrical engineer, started writing software for a power plant, and ended up building agentic AI platforms. The thread through all of it is the same: understand the constraints of the place you are building for, then ship something people actually use.
And a new adventure ahead
AI/ML & Full-Stack Engineer
ThreadChainExperienceMay 2026 - Present- Build the Gemini orchestration behind an AI design-specification generator for fashion brands, producing garment renders, artwork and measurement tables from a questionnaire.
- Generation pipeline across Vertex AI and the Gemini Developer API, selected per environment by a client factory.
- Per-call AI cost metering via OpenTelemetry counters and a PostgreSQL audit log, surfaced through billing-summary endpoints.
- Centralised retry layer with equal-jitter exponential backoff on 503, 429, quota and network-reset errors, stabilising generation across ~20 Celery tasks.
- Stripe credit billing with a 402-triggered paywall intercepted centrally in the API client.
- Observability across FastAPI, Celery and SQLAlchemy into Prometheus, Grafana and Jaeger.
- Full-stack delivery on a Next.js 15 / React 19 frontend and a FastAPI backend of 63 models and 62 API routers.
AI/ML Engineer
Metcore Digital TechnologiesExperienceSep 2025 - Present- AI/ML engineering across all product lines - telecom AIOps, fintech, construction, education and enterprise knowledge platforms.
- Architected LangGraph agents on AWS Bedrock AgentCore across 7 product lines, with HMAC-signed identity tokens and SigV4 signing enforcing tenant authorisation at the tool-call boundary.
- Shipped the same agents across three delivery surfaces: speech-to-speech voice on LiveKit with Amazon Nova Sonic, a per-tenant API-key REST endpoint streaming over SSE, and an embeddable JavaScript widget.
- Built an AIOps operator console for a national telecom, integrating six monitoring systems behind a connector SDK, with anomaly detection, forecasting, change-point detection and root-cause ranking.
- Enterprise RAG platform on AWS S3 Vectors with Bedrock Titan v2 embeddings, over a 35-table schema with queued web-crawl ingestion.
- Bedrock agent exposing 31 custom tools behind a circuit breaker and a custom MCP server.
AI & Full-Stack Engineer (Contract)
Legal Action MediaExperienceJul 2025 - May 2026- Owned the AI layer across the product suite of a US personal-injury law firm, remote from Lagos.
- LangGraph/LangChain pipelines with PostgreSQL checkpointing for resumable claim-intake and document-processing workflows.
- FastAPI service orchestrating OpenAI Agents SDK agents with LiteLLM multi-provider routing and live WebSocket progress streaming.
- Full-stack delivery with the platform team on a Next.js 16 claims-management system, integrating Twilio Voice/Video, Nylas, DocuSeal and Stripe.
Industrial Engineering Intern (NYSC)
Egbin Power PlcExperienceJul 2024 - Jul 2025- One-year national service placement at one of Nigeria's largest thermal power stations, building internal tooling in Power Apps, Python and TypeScript.
- Cut OPEX report preparation from one week to two days with a desktop ticket-merger application.
- Eliminated a backlog of dozens of unattended tickets by replacing manual triage with automated routing across a 430-bed estate serving 500+ staff.
- Shipped a budget-management application adopted by every department, administered by finance.
B.Eng. Electrical/Electronics Engineering
Federal University of Technology, MinnaEducation2018 - 2023- Second Class Upper. Led the campus Google Developer Student Clubs chapter.
- Google Solution Challenge - Minna Champions (2023), leading an AI-powered waste management platform aligned to SDG 12.
- PAYAZA Hackathon 3.0 - 4th place as team lead, building an AI solution for MSME financial inclusion.
Machine Learning & AI Programming Nanodegrees
UdacityEducation2023- Machine Learning Fundamentals and AI Programming with Python - deep learning, computer vision and model deployment.
And a new adventure ahead
Featured Work
LLM infrastructure, autonomous agents, and retrieval-augmented generation systems spanning data pipelines, vector search, production APIs, and frontends.
Multi-Tenant Agentic AI Platform
An agent platform where each product gets its own agent, but none of them hard-code their tools. Agents discover tool catalogues at runtime over MCP (streamable HTTP), generated from each backend's vendored OpenAPI contract, with nightly contract-drift detection in CI. Tenant authorisation is enforced at the tool-call boundary itself using HMAC-signed per-request identity tokens and AWS SigV4 request signing.
Enterprise RAG Platform
A retrieval platform where every tenant gets an isolated vector index. Documents and crawled pages are chunked and embedded with Bedrock Titan v2 (1024-dimension, cosine), then written to AWS S3 Vectors rather than a conventional vector database. Ingestion runs asynchronously through BullMQ-queued Crawlee crawls and PDF processing, across a 7-service Turborepo backed by a 35-table Drizzle/PostgreSQL schema.
AIOps Operator Console
Telecom operations teams watch six different monitoring systems - Grafana, StableNet, Azure Monitor, Elastic, Metabase and Wavefront. This console puts them behind a single connector SDK with a common contract, then runs a Python/FastAPI analytics service over the unified telemetry: anomaly detection (isolation forest, random cut forest and seasonal methods), forecasting, change-point detection, log-template mining, correlation and root-cause ranking, over a 64-table schema.
Per-Call LLM Cost Instrumentation
Most LLM cost tracking multiplies total tokens by one number, which is wrong the moment a model prices image output differently from text, or bills thinking tokens separately. This system instruments every model call with OpenTelemetry counters, prices each token class against a maintained per-model rate table, and persists the result to a PostgreSQL audit log exposed through billing-summary endpoints. It sits alongside a centralised retry layer using equal-jitter exponential backoff on 503, 429, quota-exhaustion and network-reset errors.
Ensemble Sports Prediction Platform
A Flask ML microservice combining three models - Poisson regression for score lines, an XGBoost classifier for outcomes, and an Elo baseline - in a weighted ensemble (0.35 / 0.45 / 0.20), with Monte Carlo simulation for match outcomes. GPT-4o generates commentary behind a regional compliance filter with rate-limit retries and token-usage tracking, served through tiered Stripe subscriptions on a 5-service Docker stack.
Technical
Stack
The technologies I reach for most often when building LLM infrastructure, agents, and retrieval-augmented generation systems.
Learning fast.
GenAI & LLM Infrastructure
10 technologies
The agent, orchestration and model stack I use to build LLM-backed products that survive production.
Machine Learning
10 technologies
Classical ML and deep learning, from ensembles and anomaly detection through to computer vision.
Languages & Backend
9 technologies
The languages and services I use to build systems end to end.
Data & Infrastructure
10 technologies
The storage, retrieval and deployment tooling behind the systems I ship.
Talk to my
AI Twin
Checking whether the live assistant is available right now. You can still open the chat panel while the status loads.
Best for recruiter summaries, architecture deep-dives, LLM infrastructure and agent discussions, and guided project walkthroughs.
Get in Touch
Open to Software Engineering, AI/ML Engineering, and LLM infrastructure opportunities, as well as thoughtful collaborations.
Emmanuel Mayowa Samuel
AI/ML Engineer
Book a call
A 30-minute conversation about the role, the work I've done, and whether it's a fit.