ym Yuguda.
Résumé

Machine Learning Engineer · Production LLM Systems · Lagos, Nigeria

Yuguda Muhammed.

I build models that survive production — LLM reasoning agents, retrieval over messy real data, and the serving and evaluation layers that keep them honest once real users show up.

Selected work Get in touch
01 — Work

Selected work

Production ML systems, and the research builds behind them.

All projects ↓
Autonomous multi-step LLM reasoning over government solicitations
Adept Engineering Solutions · In production
01

Proposal Generation Engine ↗

An LLM agent that ingests raw solicitation PDFs, structures the unstructured, reasons over it in stages, and returns scored, schema-validated output through a FastAPI service. Output quality moved from 30% to 90%+ under a repeatable evaluation harness.

  • LangGraph
  • Multi-step agents
  • Structured output
  • Evaluation harness
  • FastAPI
  • AWS
vLLM inference server logs serving chat completion requests
02

FL-04 Orchestrator ↗

A draft → red-team → revise documentation pipeline on a locally served open-weight model, and the design work to turn that fixed workflow into an agent that decides which steps a change actually needs.

  • vLLM
  • Open-weight models
  • Agent design
  • Python
CampusPQ OpenAPI documentation showing service endpoints
03

CampusPQ ↗

Multimodal learning platform — PDF, image, audio and video pipelines with LLM analysis and vector search behind them. Scaled to 5k+ processed questions at a 98% success rate.

  • Multimodal
  • VectorDB
  • FastAPI
  • AWS
Self-hostable memory server for AI clients
Open source · Published on PyPI
04

DevMemory ↗

A multi-user MCP memory server running fully locally with no external API dependency, exposing reusable tooling across Claude, Cursor and Windsurf. JWT and API-key auth, tiered quotas, migrations, monitoring dashboard, CI.

  • MCP
  • AI infrastructure
  • SQLAlchemy
  • Docker
A vision model on WhatsApp, running on open weights
Gemma-4 · Meta Cloud API
05

Eeva ↗

A multimodal assistant that analyses product images and generates marketing copy over WhatsApp, driven by a stateful message router for iterative refinement — built on an open-weight vision model rather than a proprietary API.

  • Gemma-4
  • Vision-language
  • State machine
  • FastAPI
Is this receipt real, or generated?
EfficientNet-B0 + BERT + TrOCR
06

Receipt Classifier ↗

A hybrid image–text model for detecting synthetic receipts: EfficientNet-B0 over pixels, BERT over text pulled out with TrOCR, fused into one classification head.

  • PyTorch
  • Multimodal fusion
  • OCR
  • Computer vision
5+
Years building and deploying ML systems in Python
30 → 90%
Output quality lift on the production reasoning system
5k+
Questions processed at a 98% success rate on CampusPQ
~100
Concurrent users held inside low-latency serving targets
Gradient descent · loss surfaceθ ← θ − η∇J(θ)
init θ₀ min
02 — Capabilities

What I do

The full lifecycle — data and modelling through serving, evaluation and monitoring.

01 / Reasoning

LLM & agent systems

Reasoning systems that survive contact with messy input and real users.

  • Multi-step agents — LangChain, LangGraph, CrewAI
  • RAG, embeddings and semantic retrieval
  • Prompt optimisation and output validation
  • Open-weight models served locally (Gemma, vLLM)
02 / Modelling

Training & evaluation

Fitting the model, and — the harder part — knowing whether it actually works.

  • PyTorch, Hugging Face Transformers, scikit-learn
  • Transformers, LSTMs, CNNs, classical ML
  • Evaluation benchmarks and repeatable scoring
  • Computer vision, OCR and multimodal fusion
03 / Infrastructure

ML infrastructure

The unglamorous layer that decides whether a model ever reaches anyone.

  • FastAPI inference APIs, JSON contracts, OpenAPI
  • Qdrant, ChromaDB, pgvector, PostgreSQL, Redis
  • Docker, Kubernetes, GitHub Actions CI/CD, AWS & GCP
  • Latency monitoring, model and dataset versioning
03 — Track record

Experience

Download CV ↗
Jul 2024 — Present

Lead AI Engineer

Adept Engineering Solutions · USA · Remote
  • Lead design and delivery of an end-to-end production LLM system applying multi-step reasoning agents over structured data, served via FastAPI on AWS.
  • Drove output quality from 30% to 90%+ through iterative model evaluation, prompt optimisation and structured output validation.
  • Built ingestion, preprocessing, embedding and semantic-retrieval pipelines over noisy real-world data, with explicit handling for edge cases and failure modes.
  • Engineered for ~100 concurrent users: Redis caching, horizontal scaling, latency monitoring.
Sep 2024 — Oct 2024

LLM Agent Trainer & Evaluator

Highbrow Technology · Contract · Remote
  • Defined evaluation criteria and quality-assured structured LLM outputs for correctness, reasoning quality and schema compliance.
  • Applied prompt engineering and data refinement to improve model reasoning accuracy.
May 2022 — Jan 2025

AI / Backend Engineer

Bitsaac Inc. · Remote · FinTech / E-commerce
  • Built RAG-powered analysis pipelines and semantic retrieval systems using LLMs as a reasoning layer over structured data.
  • Improved vector retrieval performance by 30% via embedding-model evaluation, dataset-quality work and ChromaDB / Qdrant tuning.
  • Shipped RESTful JSON APIs for model operationalisation, with fallback behaviour for provider failure.
2018 — 2024

B.Sc. Physics

Ahmadu Bello University, Zaria
  • A physics background underpins the quantitative rigour, systems thinking and methodical documentation behind applied ML work.
04 — Archive

More projects

Research builds, applied experiments, and the systems work underneath them.

GitHub ↗
Tagora Visual similarity search over CLIP embeddings with FAISS indexing, exposed as a REST API. CLIP · FAISS Next-Word Prediction Encoder-only transformer written from scratch, a GPT-style decoder, and a fine-tuned GPT-2, compared on one corpus. Transformers Hybrid LSTM–LR Forecasting Stock forecasting combining an LSTM with linear regression to separate non-linear signal from trend. Time series RAG PDF Parser Concurrent document processing with embeddings and semantic query over extracted content. RAG PocketClinic Triage assistant for resource-constrained clinics — symptom analysis against WHO protocols over SMS and voice. Agents · Twilio CrewAI Trip Planner Multi-agent planning system with specialised agents collaborating on a shared task graph. Multi-agent Content Repurposer Production service ingesting blog content and orchestrating AI repurposing through a Celery queue. Celery · LLM Spam Classification Text classification pipeline benchmarking NLP feature extraction against several classical algorithms. NLP Loan Approval Prediction Tabular classification with exploratory analysis of the factors actually driving approval decisions. scikit-learn Project Edix Systems work — school records anchored to the Cardano blockchain so certificates verify without exposing student data. Systems design

Let's build something that ships.

Open to machine learning and AI engineering roles, remote or hybrid.

Email me Book a slot ↗ LinkedIn ↗

Prefer email? elemenx93@gmail.com