Open to senior AI/ML roles & graduate research

Joseph Azumbil Agere · AI/ML Lead · Accra, Ghana

Building AI systemsthat actually ship.

AI/ML Lead at 4th-IR. I design and deliver multi-agent platforms, hybrid RAG, and production LLM systems — from prompt pipelines to evaluation and deployment.

View selected workGet in touch
Scroll
01About

I lead an AI team and ship systems end to end.

I lead an AI engineering team and ship AI systems end to end.

Currently AI/ML Lead at 4th-IR, where my recent focus is LLMs and multi-agent platforms built with LangGraph and MCP, hybrid retrieval, and LoRA fine-tuning. I'm equally comfortable across NLP, classical ML, and the MLOps needed to get models into production and keep them reliable.

Beyond the modeling, I set the standards a team ships against — code review, model evaluation, prompt pipelines, and deployment — and I'm the lead technical contact on client engagements, from scoping to delivery. I write and communicate in English at a C2 level.

02Experience

Three years shipping in production.

From software delivery to leading an AI engineering team — a steady move toward harder, higher-stakes systems.

  1. AI/ML Lead · 4th-IR

    Jun 2024 — Present
    • Delivering Twynity, a virtual-workforce platform where users create digital twins — each with its own persona, skills, and knowledge base — that handle tasks autonomously. Built the agent layer with LangGraph for orchestration and MCP servers for tool access.
    • Built production agents for MiFID II compliance checking, KYC automation, and obligation extraction, now in use by enterprise clients.
    • Designed and iterated prompt pipelines for production LLMs — multi-step fallback chains and structured output — for reliable, high-quality responses at scale.
    • Established the team's model-evaluation standards: quality benchmarks, safety checks, and human-review workflows that gate AI responses before client deployment.
    • Designed an Autonomous Data Modeler that turns a business spec into logical data models, then maps to existing warehouse tables or proposes new schemas — cutting data-modeling time on a recent engagement from weeks to hours.
    • Built hybrid RAG pipelines (dense + sparse retrieval with feedback memory) for cases where vector search missed the right chunks; used across several deployments.
    • Built reusable MCP servers so agent tools — search, retrieval, business APIs — are shared across client projects, cutting integration time per engagement.
    • Lead technical contact on engagements. Set team standards and mentor two junior AI engineers.

  2. AI/ML Engineer · Inngen

    Nov 2023 — Jun 2024
    • Built Frank, a regulatory chatbot used daily by staff at the Frankfurt Health Department for compliance lookups — LangChain, Pinecone, and FastAPI with RAG over their policy documents.
    • Built NLP pipelines for enterprise document review, cutting review time by around 70% on the main client workflow.
    • Fine-tuned and deployed domain-specific LLMs on AWS SageMaker.
    • Designed multi-step prompt pipelines with fallback handling for cases where the primary path failed.

  3. Software Developer · AfroTechnologies

    Jun 2023 — Nov 2023
    • Built a Senior High School admissions system that replaced in-person enrollment for thousands of students, cutting administration and travel time by around 60%.
    • Contributed to the deployment of a Learning Management System used by several schools for virtual teaching.
03Selected Work

Systems built for real clients.

Most of this is confidential client work, so it's presented as case studies — the problem, the approach, and the outcome — rather than open repositories.

Twynity

Client work

A virtual-workforce platform of autonomous digital twins.

Built the agent layer with LangGraph for orchestration and MCP servers for tool access, so each twin reasons, calls tools, and acts on its own.

Impact

Digital twins that handle real tasks autonomously across the platform.

LangGraphMCPMulti-agentLLMs

Compliance & KYC Agents

Client work

Production agents for MiFID II checks, KYC, and obligation extraction.

Designed prompt pipelines with multi-step fallback chains and structured output, gated by explicit evaluation and human-review standards before deployment.

Impact

In production with enterprise clients.

LLMsStructured outputPrompt pipelinesEvaluation

Autonomous Data Modeler

Client work

From a business spec to logical data models — automatically.

Built a system that takes a business spec, generates logical data models, and either maps to existing warehouse tables or proposes new schemas.

Impact

Cut data-modeling time on a recent engagement from weeks to hours.

LLMsData modelingWarehousesSchema design

Hybrid RAG Pipelines

Client work

Retrieval that finds the chunks plain vector search misses.

Combined dense and sparse retrieval with a feedback-memory loop so the pipeline learns from what it gets wrong.

Impact

Adopted across several client deployments.

Hybrid retrievalPineconeWeaviateFeedback memory

Frank

Client work

A regulatory chatbot for the Frankfurt Health Department.

Built RAG over their policy corpus with LangChain, Pinecone, and FastAPI, plus prompt pipelines with fallback handling.

Impact

Used daily by department staff.

LangChainPineconeFastAPIRAG
04Capabilities

The stack I build with.

Depth across the modern AI engineering stack — from agents and retrieval down to the MLOps that keeps it all running.

LLMs & Generative AI

  • OpenAI
  • Anthropic Claude
  • Mistral
  • LoRA / domain fine-tuning
  • Prompt engineering
  • Model output evaluation

Agents

  • LangGraph
  • LangChain
  • LlamaIndex
  • MCP servers
  • Multi-agent orchestration

RAG & Retrieval

  • Hybrid retrieval
  • Pinecone
  • Weaviate
  • Embedding pipelines
  • Feedback memory

ML Frameworks

  • PyTorch
  • TensorFlow
  • scikit-learn
  • Model evaluation

MLOps & Cloud

  • AWS
  • AWS SageMaker
  • Azure
  • Docker
  • CI/CD
  • Terraform

Backend & APIs

  • FastAPI
  • Django
  • Flask
  • GraphQL
  • REST
  • PostgreSQL
  • MongoDB

Languages

  • Python
  • JavaScript
  • PHP
  • C++
05Education & Research

Foundations &where I'm headed.

A computer-science grounding, an active research interest, and a habit of teaching what I learn.

B.Sc. Computer Science

2019 — 2023

University for Development Studies · Ghana

Coursework
  • Artificial Intelligence
  • Data Structures & Algorithms
  • Software Engineering
  • Cloud Computing
  • Linear Algebra
  • Algorithm Design & Analysis
Research interests
  • Multi-agent LLM systems
  • Retrieval-augmented generation
  • Evaluation & safety of generative models
  • Efficient fine-tuning (LoRA)
Beyond work

IndabaX Ghana 2025

Ghana Data Science Summit · Ashesi University

Attended workshops on multimodal LLMs and LoRA fine-tuning.

Mentor — Teens in AI (Africa)

Since 2024

Mentoring the next generation of African AI builders.

First Runner-Up — FinTech Hackathon 1.0

Trestle Academy Ghana · 2022

Co-built a virtual-card platform with AI fraud prevention for cross-border transactions.

06Contact

Let's build something worth shipping.

Open to senior AI/ML roles, client engagements, and graduate research conversations. The fastest way to reach me is email.

agerejoseph448@gmail.comLinkedInGitHub
Accra, Ghana(+233) 541 433 448English · C2