Work that made it into production

The Challenge

Property portals are excellent at showing what is for sale, and almost silent on the question that actually matters: is this one right for me? Every listing looks the same — a price, some photos, a glowing blurb — and the work of judging it against your budget, your commute, your must-haves and your dealbreakers is left entirely to the buyer, spread across a dozen browser tabs. The tools that did exist were built for investors, agents and sourcers: yields, spreadsheets and jargon. Nobody had built anything for the people actually buying homes — first-time buyers, growing families, anyone simply trying to find somewhere to live.

Our Solution

We built Nookaly to answer that one question, as an app people carry with them. Every property gets a single score out of 100, measured against four things the buyer decides for themselves: getting around, the home itself, the area, and outdoor space. Move the sliders and every score updates — someone who cares most about a short commute sees a different ranking to someone who wants a garden. Affordability is a filter, not a factor. A property outside the budget never appears at all, because a home you cannot afford is not a compromise worth scoring. Search, shortlists, favourites and a personalised daily feed all live in the iPhone and Android apps, so the answer is there on the sofa, on the train, and standing outside the front door at a viewing. A companion browser extension carries the same scores onto Rightmove and Zoopla for anyone who prefers to start on a desktop. Each listing is enriched automatically with the context buyers would otherwise chase themselves: real journey times, planning applications nearby, air quality, safety and schools — around 20 separate metrics behind every score.

Results

• Turns a dozen open browser tabs into one number a buyer can act on
• Buyers get a ranked answer in their pocket instead of a shortlist of maybes across a dozen tabs
• Scores reflect each person's own priorities rather than a generic ranking, and update instantly when those priorities change
• Live on the App Store and Google Play, with a companion browser extension on the Chrome Web Store
• Demonstrates Byte Insights taking a product from idea to launch, not only automating an existing process

The Challenge

AI code review tools are easy to demonstrate and hard to adopt. Most sit in a separate dashboard, review one file at a time so they miss anything that spans a change, and produce so much low-value commentary that teams stop reading it. The ones that do run in a pipeline tend to fail the build when the model is unavailable, which is the fastest way to get an integration removed. The question was whether an AI reviewer could earn a permanent place in a real team's workflow rather than being switched off after a fortnight.

Our Solution

We built it as an Azure DevOps pipeline task, so it runs on the branch policy build teams already have, and posts to the pull request rather than anywhere new. The review is holistic by default: the model reasons across every changed file at once, together with the PR title and description, linked work items and their acceptance criteria, existing human comments, and the team's own coding standards. That is what lets it catch cross-file breakage and changes that do not match their stated intent. The adoption problems are handled deliberately. Findings are fingerprinted, so a re-run updates the summary in place rather than posting everything again, and findings that no longer apply are closed automatically. Each is severity-rated with a minimum threshold, so a team can turn the noise down instead of off. An optional verification pass asks the model to confirm or drop each candidate before it is posted. Large pull requests are split, reviewed in parts and synthesised into one summary. Most importantly it never blocks the build. An API error or unparseable response reports SucceededWithIssues and warns; it does not fail the pipeline. The reviewer is advisory, and behaves like it.

Results

• Published on the Visual Studio Marketplace under the Byte Insights publisher, with over 100 installations
• Supports seven model providers — OpenAI, OpenAI-compatible, Azure OpenAI, Gemini, Vertex AI, Anthropic and local models via Ollama
• Reviews the whole pull request at once, so cross-file breakage is caught rather than missed between files
• Idempotent re-runs: nothing is double-posted and resolved findings close themselves
• Advisory by design — a model outage warns rather than failing the build
• Maintained through two major versions since March 2025

The Challenge

Answering a single support ticket often means moving between the Azure portal, the Microsoft 365 admin centres, Microsoft Graph and the command line — and knowing which of them Microsoft has put each setting behind. The information is available; finding and combining it is the work. The obvious AI answer is to give an assistant broad standing access, which is precisely what no one administering a tenancy will accept.

Our Solution

It is a Model Context Protocol server, and deliberately a thin one: a wrapper over the Azure CLI and Microsoft Graph exposing two controlled tools — one for Azure, one for Microsoft 365 — that a supported AI client can call on the engineer's behalf. Microsoft already provides the interfaces; what was missing was a safe way to let an assistant use them. The design point is what it cannot do. It runs as the signed-in account and Microsoft continues to enforce that account's permissions, so the assistant cannot grant itself anything its operator does not already have. Every call is validated by the server before it runs. An engineer can ask for a user's account, group memberships, assigned licences and managed devices in one sentence, and the assistant works out which calls that needs. It ships as a Docker image containing the server, Python and the Azure CLI, so there is nothing to install and the AI client starts and stops it automatically.

Results

• One request replaces a manual trip across several portals and admin centres
• Deliberately thin — it wraps interfaces Microsoft already provides rather than reimplementing them
• Runs as the signed-in engineer, so existing Microsoft permissions still apply and cannot be escalated
• Works with Cursor, Google Antigravity, OpenCode and OpenAI Codex
• Ships as a self-contained Docker image — no local Python or Azure CLI setup
• Open source and publicly available
• Evidence of the Microsoft estate work that stalls many AI projects

The Challenge

A financial services firm operating across multiple global markets needed to produce consistent user guides in 6 different languages for their products and services. Their existing process was time-consuming, requiring manual drafting and translation which caused delays in documentation and created inconsistencies across languages.

Our Solution

We built a secure, internal web platform that allowed the client to upload reference materials and automatically generate comprehensive user guides in 6 languages. LangChain orchestrated generation through Azure OpenAI, with prompt and terminology controls holding financial language accurate and consistent across every language, and a custom frontend for uploading reference material and configuring the output. The hard part was not generation. It was making the output consistent enough, and integrated tightly enough with how the firm actually produced documentation, that it replaced the existing process rather than sitting alongside it.

Results

• Reduced documentation time from weeks to hours per guide
• Saved hundreds of hours of staff time monthly across all departments
• Achieved consistent terminology and voice across all 6 languages
• Removed external translation agencies from the process — draft financial documentation no longer left the firm
• Still in daily production use

The Challenge

A personal and business finance coaching firm was spending hundreds of hours manually processing client financial statements. When new clients signed up, staff had to manually create folders, send welcome emails, wait for statement uploads, and then tediously enter transaction data line-by-line into spreadsheets for client review - a critical but time-consuming part of their financial literacy journey.

Our Solution

We developed a comprehensive automation system that streamlined the entire client onboarding and data processing workflow. The system integrated multiple components: 1. A Python application that connected to Typeform to capture client signups, automatically created personalised Google Drive folders, and sent customised welcome emails with upload instructions 2. A monitoring service that detected when clients uploaded financial statements to their Drive folders 3. A secure processing pipeline built on Amazon Bedrock, running OCR extraction over the uploaded statements — with a separate pipeline tuned per bank, because statement formats vary far too much for one generic pass to be reliable 4. Integration with Make.com to automatically populate Google Sheets with the extracted transaction data 5. Automated email notifications to clients when their statements were successfully processed

Results

• Reduced manual data entry from hours to minutes per client
• Saved hundreds of staff hours monthly across the entire client base
• Improved client experience with faster turnaround times on statement processing
• Automatic deletion of sensitive information within 2 hours
• Still in daily production use
• Increased capacity to take on more clients without additional administrative staff

The Challenge

An IT managed service provider was struggling with the initial processing of support emails. Their technical staff was spending valuable time manually reading, categorising, and responding to incoming support requests before they could even begin troubleshooting. Additionally, spam emails were creating unnecessary noise in their support queue, further reducing efficiency.

Our Solution

We built an intelligent email triage system using Azure OpenAI that sits between the support inbox and the ticketing system, providing an additional layer of security and efficiency. The system: 1. Automatically analyses incoming support emails and categorises them by issue type, urgency, and required expertise 2. Filters out spam and non-support emails, preventing them from entering the ticketing system 3. Generates tailored troubleshooting suggestions for IT engineers based on the specific issue detected 4. Creates personalised response emails to clients acknowledging their issue and providing initial self-help guidance where appropriate 5. Prepares the categorised information for seamless handoff to the ticketing system

Results

• Reduced initial response time from hours to minutes
• Decreased ticket categorisation errors by 65%
• Eliminated nearly 100% of spam tickets from entering the system
• Provided IT engineers with relevant troubleshooting suggestions, improving first-time fix rates
• Enhanced client satisfaction with prompt, personalised responses
• Still in daily production use