Primary focus
AI Engineer
LLM applications and Python backend services: structured outputs, tool calling, RAG, and event-driven pipelines on AWS and Microsoft 365.
AI Engineer CV (PDF)Anslem Akadu
AI Engineer · Applied AI · LLM & Backend Systems
Building practical AI systems from idea to production.
I build LLM powered backend systems in Python: event-driven services on AWS and Microsoft 365 that classify, extract, retrieve, summarise and act on real business data, with structured outputs, tests and deployment I own. Before that, I fine-tuned open weight models and built forecasting systems in health tech.
01 / About
The work is bigger than the model.
I enjoy bringing AI closer to real world applications. That means getting hands on with the technical problem, choosing a sensible approach, and staying involved through backend integration, testing, deployment, and iteration.
Most of my recent work lives inside Microsoft 365 and AWS: services that react to email, chat and meeting events, ask a model a clearly scoped question, check the answer, and deliver something a person can act on.
How I build
- Treat the model as one component: validate its output and plan for it being wrong.
- Keep private data out of prompts unless it needs to be there.
- Make every step safe to retry: queues and idempotent writes, not hope.
- Weigh latency and cost before reaching for the biggest model.
- Test the unglamorous paths; that is where incidents start.
Coffee, code, deploy, panic, repeat.
Specialist depth
Machine Learning Engineer
Fine-tuning and serving open weight models (QLoRA, quantized GGUF), forecasting with XGBoost and Prophet, and inference tuned for cost.
ML Engineer CV (PDF)02 / Experience
Engineering across the stack.
Current work first: AI services on Microsoft 365 and AWS. Earlier: fine-tuning, inference and forecasting in health tech.
AI/ML Engineer
20four7VA · Remote staffing company · Remote
I design, build, test and deploy the company’s AI services in Python on AWS serverless, integrated with Microsoft 365. These are private employment systems, so the code is not public.
- Executive email intelligence. A multiuser Python service on AWS SAM (Lambda, SQS, DynamoDB) that ingests Microsoft Graph mail and calendar webhooks, triages each email by priority (an LLM rubric plus deterministic rules), and sends urgent alerts and Pydantic validated daily and weekly briefs to Teams.
- Teams AI assistant. A streaming tool calling agent (18 tools) over Graph mail, calendar and meeting transcripts that drafts replies and books meetings with delegated user tokens, and sends nothing until the user approves it on an Adaptive Card. Briefs are also exposed in Microsoft 365 Copilot.
- Teams compliance monitoring. A Graph change notification pipeline (API Gateway, SQS, Lambda) with a trigger filter of 146 phrases, conversation context retrieval, redaction of 17 PII and secret types before inference, and strict JSON schema classification against 11 policies. Flagged cases alert the right manager in Teams.
- Lead and meeting intelligence. GoHighLevel and Zoom webhook pipelines (SQS, Lambda) for multistage LLM extraction and classification, lead enrichment and scoring with DataForSEO and SerpAPI, and a RAG pipeline that retrieves similar past job descriptions with OpenAI embeddings so new JD drafts follow what has worked before.
- Business continuity monitoring. A scheduled Lambda that matches GDACS disaster events and WMO weather alerts to around 200 geocoded staff locations (haversine distance with tiered 50/100 km risk levels, Shapely for alert polygons) and sends deduplicated Teams alerts that escalate when the risk rises.
- Reliability and testing. Idempotent DynamoDB conditional writes, Graph subscription renewal, dead letter queues with CloudWatch alarms, per user data isolation, and nearly 300 automated pytest tests using moto and fakes.
Engineering decisions worth asking me about
- Redaction runs before the model call, and every compliance verdict must match a strict schema. Failures go back to the queue and, after retries, to a dead letter queue instead of being guessed.
- The assistant can draft but not send: sending needs an Adaptive Card click, and each action is claimed once so a double click cannot send twice.
- A small model handles high volume triage and a larger one writes briefings, with token usage metered per user.
- Graph subscriptions expire, so a scheduled reconciler renews them and recreates any that have disappeared.
- The JD generator retrieves the closest past job descriptions and uses them as style references, not as facts to copy.
- Microsoft 365 Copilot reads the briefings through a declarative agent backed by an OpenAPI plugin.
ML/AI Engineer
LifeBank (Nigeria) Ltd · Health tech & medical logistics
- Fine-tuned Gemma 3 1B with QLoRA (PEFT) for domain specific hospital finance tasks: instruction alignment and domain adaptation.
- Migrated the assistant from Gemma to Qwen, cutting inference latency by 20%.
- Automated AWS EC2 GPU training environment setup with Bash, reducing setup time by 40%.
Data Scientist · Financial Systems
LifeBank (Nigeria) Ltd
- Built XGBoost and Prophet models forecasting hospital profitability, liquidity, leverage and efficiency, with hospital size and state as regressors for revenue and cash flow forecasts.
- Designed a hybrid forecast (80% naive baseline, 20% ML) to stay robust on sparse data, surfacing early warning signals of negative cash flow.
Data Scientist · Healthcare Forecasting
LifeBank (Nigeria) Ltd
- Built time series forecasting pipelines for blood, oxygen and medical stock demand, reducing stockout risk by 20%.
- Developed predictive maintenance models for medical equipment and a forecasting engine with scenario simulations for supply planning.
03 / Selected work
Built to be used.
Public code you can inspect. My employment systems are private, so this is where to read how I write software.
01
Open source · 2025
LLM FastAPI Boilerplate
A small, reproducible FastAPI service for serving quantized GGUF language models on CPU.
Pydantic request and response schemas, llama-cpp-python inference with
configurable sampling, and /generate, /chat (with
conversation history) and /health endpoints, packaged with Docker
and deployed to Hugging Face Spaces. I iterated across Phi, Mistral 7B,
Qwen2.5 and TinyLlama quantizations to fit CPU only hosting.
02
Open source · 2025
ElevatrAI
A career guidance web app that compares a resume with seven tech roles, shows the skill gap and builds a learning roadmap.
Parses PDF and DOCX resumes, extracts skills against a JSON taxonomy of 56 skills, ranks role fit with cosine similarity over skill vectors (scikit-learn), and assembles a plan in three phases from curated resources. A Flask app deployed with Gunicorn on Render.
Free tier hosting: the first load can take up to a minute.
04 / Toolkit
A practical technical toolkit.
Applied AI and backend delivery come first, supported by machine learning foundations and deployment experience.
01
Applied AI & LLM Engineering
- OpenAI and Anthropic APIs
- Structured outputs (JSON Schema, Pydantic)
- Tool calling and streaming
- Retrieval augmented generation (RAG)
- Embeddings and similarity search
- PII and secret redaction
- QLoRA and PEFT fine-tuning
- Hugging Face Transformers
- Quantized GGUF inference (llama-cpp-python)
02
Backend & AI Systems Engineering
- Python, FastAPI, and Flask
- REST APIs and webhooks
- Pydantic data contracts
- Event-driven queue workers
- pytest and moto
- SQL and Bash
03
Machine Learning & Data
- PyTorch and scikit-learn
- XGBoost and Prophet
- Time series forecasting
- Feature engineering and model evaluation
- SentenceTransformers and embeddings
- pandas and NumPy
- Geospatial matching (haversine, Shapely)
04
Cloud, Deployment & Reliability
- AWS Lambda, API Gateway, and SQS
- DynamoDB, EventBridge, and CloudWatch
- AWS EC2 (GPU)
- AWS SAM (infrastructure as code)
- Docker and Redis
- Linux, Git, and GitHub Actions
- Idempotency and dead letter queues
- Claude Code and Codex
05
Microsoft 365 & Integrations
- Microsoft Graph (mail, calendar, Teams, transcripts)
- Graph change notifications
- MSAL (app only and delegated)
- Teams bots and Adaptive Cards
- Microsoft 365 Copilot declarative agents
- Zoom, GoHighLevel, DataForSEO, and SerpAPI
05 / Contact
Let’s build something useful.
If you’re hiring for an AI Engineer, Applied AI or ML Engineer role, I’d be glad to talk about the problem your team is solving.
anslemakadu1@gmail.com