Work
Real systems.Clear boundaries.
A selection of working projects and the engineering patterns behind them. The common thread is simple: start with the information and workflow, then choose the smallest useful system.
Selected systems
Built, not invented for a portfolio.
These are public examples. They show the kind of problems and architecture I work with; they are not disguised client case studies.
DocSumm
Problem. Useful information is buried in PDFs, DOCX files, HTML, EPUBs and meeting recordings.
Approach. Extract the material locally, prepare it for a model, and produce a concise summary and key points without making the source material a public web upload.
Architecture. Python ingestion and processing, designed to sit beside a local model such as Ollama.
02 · AI DATA PIPELINEDocument Labs
Problem. Before a document reaches an AI system, you need to understand what is in it and how it will be represented.
Approach. Browser-based tools demonstrate three mundane but important steps: detect obvious sensitive data, estimate prompt size, and split material into retrieval-sized windows.
Architecture. Client-side JavaScript. The demo tools process pasted text in the browser rather than sending it to a Rametric server.
03 · WORKFLOW SOFTWAREThe Metric
Problem. Repeated job applications require the same career information to be rewritten and analysed over and over.
Approach. Keep a structured career record and use it as the source for targeted CV, application and interview workflows.
Architecture. A focused application layer around structured data, with AI used where it helps rather than as the product itself.
04 · AI-ASSISTED ANALYSISCV Analyzer
Problem. A CV can look reasonable while still missing the language and evidence a particular job description asks for.
Approach. Compare the CV with the target role and surface gaps that can be addressed before an application is submitted.
Architecture. A focused interface around structured inputs and AI-assisted analysis, separated from the paid workflow.
The architecture
Understand the data
What files exist? Who can see them? What may leave the organisation? What must be retained and auditable?
Choose the boundary
Local inference, hosted APIs, or a deliberate mix. Sensitive material stays on infrastructure you control when required.
Connect the work
Put retrieval, drafting, classification or summarisation into the actual workflow, with approval steps where they belong.
Capabilities
Local LLMs
Local model deployment, model selection, access control and separation between local data and hosted services.
Document intelligence
Extraction, chunking, retrieval and summarisation across PDF, DOCX, HTML and meeting audio.
Process automation
Short workflows that draft, classify, route or file — without pretending every decision should be delegated to a model.
Small web applications
Focused interfaces around a real process. Simple routes, explicit behaviour and as little machinery as the problem requires.
Start with one useful workflow.
Define the input, the decision or transformation, the human checkpoint and the output. Then build the smallest system that makes that workflow better.
No requirement to replace your existing stack. No need to commit to one model vendor.
Public evidence