ML Engineer and founder with 13 years building production systems. The last two have been deep in applied ML: NLP scoring pipelines, RAG systems, LLM integration, and the MLOps tooling that makes them actually run in production. I'm also completing a BSc in Mathematics at the Open University and a Cambridge Data Science Career Accelerator, because the theory matters when the work gets hard.

My background is in senior and lead engineering roles with long-term ownership of production systems across fintech, energy, defence, and SaaS.

• Distributed systems and microservice architecture, designed and evolved under real operational pressure
• Cloud-native development on AWS and Azure, including infrastructure automation and CI/CD pipelines
• Strong emphasis on reliability through TDD and BDD: unit, integration, contract, and end-to-end testing
• Production API design with strict separation between domain, application logic, and infrastructure concerns
• Languages across production systems: Python, Go, C#, SQL, JavaScript

Recent production ML work:

Gradia (Keenu.io, via Turing Innovation Catalyst) — automated IELTS writing assessment API. NLP scoring pipeline using spaCy word vectors and cosine similarity across eight subscores covering paragraph cohesion, sentence-to-document similarity, lexical richness, vocabulary sophistication, and punctuation analysis. HuggingFace BERT-based punctuation restoration model for grammatical analysis. LLM-integrated scoring with a routed client supporting OpenAI and Ollama as interchangeable backends. After applying linear regression calibration to raw LLM scores, 94% of marks landed within +/- one band and 86% within +/- half a band. Full MLOps stack: MLflow, FastAPI, Docker, AWS ECR/ECS, GitHub Actions CI/CD.

Nightshift (Cambridge Data Science Career Accelerator, employer project for NICE) — RAG pipeline for clinical code retrieval. Given a free-text clinical research question, the system parses it into typed clinical entities using Stanza biomedical NER (bc5cdr and i2b2 models), then retrieves and ranks relevant codes across SNOMED CT, ICD-10, QOF, and NHS reference sets using hybrid semantic and TF-IDF retrieval backed by SapBERT/BioBERT embeddings in sqlite-vec. MLflow evaluation against gold-standard code sets tracking F1, precision, and recall.

MarkVerify (Keenu.io) — evaluation scoring platform built in Go for collecting and human-scoring inbound IELTS test submissions, producing the ground-truth labels Gradia's MLflow experiments ran against.

Alongside production work, I've applied a broader set of ML techniques through the Cambridge programme: scikit-learn, XGBoost, Keras, TensorFlow, PyTorch, statsmodels, BERTopic, LSTM, ARIMA, anomaly detection, clustering, dimensionality reduction.

I've hired and led cross-functional engineering teams of up to 15, owned products end to end from ideation through live operation, and mentored engineers across multiple levels through pairing, code review, structured 1:1s, and company-wide talks. Most recently as Delivery Lead at BrightHR, responsible for the BrightSafe product across five countries.

Covet is a two-sided recruitment platform for the UK senior tech market. Under the hood it's an NLP and recommender system: a two-stage matching pipeline combining hard constraint filtering with embedding-based scoring using sentence transformers (BAAI/bge-large-en-v1.5) across structured candidate and role profiles.

The longer-term architecture extends the pipeline with agent-to-agent communication, where candidate and company agents negotiate constraints and conduct structured evaluation autonomously. The outputs are visible to both sides.

The stack: FastAPI, HTMX, SQLite, AWS (EC2, ECR, Cognito, CloudWatch), Docker, GitHub Actions. A standalone Scout service pulls and normalises job listings via ATS adapters (Greenhouse, Lever, Workable, Ashby).

blog.covet.digital · linkedin.com/in/david-ahearne-767614b3