Portfolio

We’re building open demonstrations of our work so you can see exactly how we approach a project. Client case studies will appear here as engagements complete.

Open demos

Working examples of how we approach AI, delivery and quality. Code links are added as each demo is published.

Client project

RAG assistant with evals and guardrails

For a global financial information company, a compliance assistant that answers from regulatory documents with clause-level citations, input and output guardrails, and an automated evaluation suite.

Result: 92% faithfulness, 94% answer relevance and a 4% hallucination rate on the evaluation set.

  • Python
  • OpenAI
  • Hybrid search
  • Reranking
  • LLM-as-judge
  • Guardrails
  • AWS
Planned

LangGraph multi-step agent

A stateful agent that plans, calls tools through MCP, retries on failure and pauses for human approval before taking actions.

What it shows: How we build agents that are traceable, testable and safe to run on real workflows.

  • LangGraph
  • MCP
  • Python
  • LangSmith
In progress

Fine-tuned small model vs large model

For a financial services group, a small language model fine-tuned with QLoRA to classify bank transactions, benchmarked head to head against a large hosted model, then deployed behind a hybrid router.

Result: 91% vs 92% accuracy, 71× faster, about $20 saved per 1,000 requests, and 97% of requests kept inside the client’s cloud.

  • QLoRA
  • Hugging Face PEFT
  • OpenAI
  • Benchmarking
  • Model routing
Client project

AI-assisted failure analysis

For an enterprise client, a k6 load testing framework with an AI stage that reads results and telemetry, groups failures by cause and writes a prioritised fix list after every run.

Result: zero manual triage, and a hidden “2xx for invalid input” API bug that standard metrics missed.

  • k6
  • Node.js
  • Claude
  • GPT-4o
  • Azure DevOps
Client project

GitHub Actions CI/CD pipeline

For a software product company, a pipeline that builds, tests and deploys a containerised application, with tests gating every merge and one-click rollback.

Result: merge to production in 15 minutes instead of 1–2 days, and 4× more releases each week.

  • GitHub Actions
  • Docker
  • Kubernetes
  • AWS
In progress

Playwright UI automation framework

A TypeScript end-to-end suite for a sample web app, using page objects, parallel cross-browser runs and HTML reports with traces and videos.

Status: framework, page objects and cross-browser setup are done; core journeys and reports are being built.

  • Playwright
  • TypeScript
  • Chromium
  • Firefox
  • WebKit
Client project

REST Assured API testing framework

For a dental practice software company, a Java API suite covering authentication, validation, contracts and negative cases, gating every merge in CI.

Result: 500+ automated tests across 80 endpoints, a 12-minute full regression, and 90+ defects caught before UI testing.

  • Java
  • REST Assured
  • TestNG
  • Allure
  • Azure DevOps

Client results

Real engagements and what changed for the team.

Case study

Compliance AI assistant: answers in seconds, with citations

Challenge: compliance teams at a global financial information company spent hours finding the right clause in long regulatory documents, and a generic chatbot was too risky.

What we did: a RAG assistant with clause-aware indexing, hybrid search and reranking, cited answers, input and output guardrails, and evals gating every release.

Result: 92% faithfulness, 94% answer relevance, 4% hallucination rate, 1% false block rate.

In progress

Fine-tuned SLM: large-model accuracy at a fraction of the cost

Challenge: a financial services group sent every bank transaction to a large hosted model, which was slow, costly at volume, and sent sensitive data outside.

What we did: fine-tuned a small open model with QLoRA, benchmarked it fairly against the large model, and shipped a hybrid router.

Result: 1 point of accuracy difference, 71× faster responses, about $20 saved per 1,000 requests, 97% of requests handled privately.

Case study

AI failure analysis: load tests that diagnose themselves

Challenge: engineers spent hours reading raw k6 output and separate dashboards to work out why requests failed.

What we did: deterministic failure classification plus an AI diagnosis stage with a rule-based fallback, run from the pipeline.

Result: zero manual triage, repeatable resilience checks, and a hidden API validation bug uncovered.

Case study

CI/CD pipeline: from 1–2 days to 15 minutes per release

Challenge: manual, laptop-based builds and releases that depended on one or two people, with broken code reaching main.

What we did: a GitHub Actions pipeline with test-gated merges, build-once Docker images, staging auto-deploys and approved production releases on AWS and Kubernetes.

Result: 15 minutes from merge to production, 100% of merges tested first, 4× more releases each week.

Case study

API-first testing: 90+ defects caught before the UI

Challenge: API bugs surfaced late through slow, brittle UI tests that rarely tried the unhappy paths.

What we did: a REST Assured framework with reusable request builders, negative cases by design and CI gating.

Result: 500+ tests across 80 endpoints, 12-minute regression runs, 90+ defects caught early.

Want a walkthrough?

We’re happy to walk you through any of these demos on a call and talk about how the same approach would work for your product.

Book a walkthrough