Products we built, and run every day

The fastest way to judge a software partner is to look at what it builds for itself. These are our own products — designed, engineered and operated by the same team that would build yours. Several run our own business.

The portfolio

LedgerLite — AI bookkeeping and invoicing for small businesses

A complete accounting application: invoicing with templates and recurring schedules, automatic overdue chasing, bank import and reconciliation, expenses, supplier records, P&L and aged-debtor reporting, and a live tax position with VAT-threshold tracking. Its standout trick is legacy paperwork — point it at a pile of historic invoices or receipts, as PDFs or photos, and OCR plus language models read them into clean, reviewable records in bulk. It runs our own books.

• Bulk import of historic invoices and receipts — OCR and LLM extraction, reviewed before saving
• Recurring invoicing with opt-in auto-send, plus automated overdue chasing by email
• Bank statement import and two-way reconciliation with auto-matching and split payments
• Tax position: VAT threshold tracking, estimated liability, set-aside and deadlines
• An AI assistant that proposes changes and never writes to the ledger without approval

Next.js · Postgres · Drizzle · self-hosted, with a desktop build from the same codebase

WhoCanFindMe — Is your business visible to AI search?

A monitoring product for the question every business is starting to ask: what do ChatGPT, Gemini, Perplexity and Claude say when someone asks about us? It runs real buyer questions against all four engines, records whether you appear, whether you are recommended, whether you are linked, and how accurately you are described — then tracks how that shifts week to week, because AI answers are not stable. Free scan, shareable report.

• Real buyer questions run against ChatGPT, Gemini, Perplexity and Claude
• Scores appearance, recommendation, citation and description accuracy
• Breaks visibility down by query type — brand, discovery, comparison, high-intent, reputation
• Shows which third-party sources the engines actually cite about your category
• Tracks the trend over time and flags what changed

Multi-engine orchestration · scheduled measurement · public shareable reports

Meeting intelligence — Private meeting notes, entirely on your own hardware

Capture a meeting from the browser, transcribe it locally, and get a structured note written to your own knowledge base — with action items pushed into your task tracker automatically. The point of the build was sovereignty: speech-to-text and the language model both run on our own machines, so no audio and no transcript ever reaches a third-party API. It is the reference implementation behind our private-AI work.

• Local speech-to-text and local language models — no cloud AI in the pipeline
• Structured notes merged with whatever you typed during the call
• Action items become tracked tasks automatically
• Semantic search across past notes for related context
• Network identity as the login — no separate account to manage

Self-hosted models · vector search · runs on our own GPU hardware

Outreach engine — Self-hosted outbound that finds, writes and sends — safely

A complete outbound platform: it discovers prospects, researches them, drafts personalised emails, and sends them under strict safety rules — with a person able to review anything at any point. Two AI passes do the writing, a drafter and then a reviewer that scores and rewrites whatever falls short of house style. Nothing reaches a recipient unless every safety condition passes, and one bad signal pauses the whole system rather than pressing on.

• AI prospecting that designs its own searches, finds a named contact, and only keeps verified addresses
• Two-pass drafting — one agent writes, a second reviews and rewrites — against an enforced house style
• An automated quality gate that blocks placeholders, spam phrasing and duplicates; clean drafts pass, the rest wait for a human
• Fail-closed sending: several independent checks must all pass, and any single failure stops the send
• Mailbox pool with rotation, staged volume ramps, and warm-up that measures real inbox-versus-junk placement
• A bounce circuit-breaker that pauses everything automatically when deliverability degrades
• Replies classified and routed into one inbox, with automatic unsubscribes and follow-ups that stop the moment someone answers

Self-hosted pipeline · mailbox health scoring · campaign dashboard and activity monitoring

AI voice nurse assistant — A conversational assistant for care settings

A voice-driven assistant built for a healthcare context, handling spoken interaction rather than a screen and a keyboard. Voice is the right interface wherever hands are busy and attention belongs on a person rather than a form — and it is also the least forgiving, which is why the clinical boundary is designed in: the assistant handles conversation and administration, and anything approaching clinical judgement routes to a qualified human.

• Natural spoken interaction rather than forms and screens
• Hard clinical-administrative boundary — no diagnostic or advisory output
• Escalation to a person designed in from the start
• Deployable privately where patient data cannot leave the premises

Speech recognition · language models · text-to-speech · private deployment options

Multi-agent research platform — Agent meshes doing continuous analysis

A production multi-agent system: specialist agents research, cross-check each other, and reach a position, with retrieval over live data feeds and a full record of how each conclusion was formed. It is where much of our agentic-workflow practice was proven — orchestration with state, retries, verification gates and audit trails are not theory for us, they are what keeps this thing honest.

• Specialist agents that research, verify and challenge one another
• Retrieval over live data feeds rather than model memory
• Verification gates before any conclusion is accepted
• Full audit trail of how each position was reached

Multi-agent orchestration · vector databases · scheduled and event-driven runs

This website's blog pipeline — An agent that writes, checks itself, and publishes

The blog on this site is written by an autonomous pipeline: it ingests industry sources weekly, researches a topic, drafts, rewrites for voice, then puts the result through a quality gate that rejects any statistic it cannot trace to an approved fact or a cited source. Posts that fail three times are skipped rather than forced. You are reading its output — and this paragraph is a fair demonstration of the honesty rule it runs under.

• Weekly ingestion of industry sources, scored for relevance
• A quality gate that mechanically rejects unverifiable numbers
• Skips publication rather than shipping a weak post
• A kill switch and a hard volume cap, both enforced in code

Scheduled agent pipeline · deterministic QA gate · automated publishing

Want something like this built for you?

Every product here started as a scoped first release, not a twelve-month programme. Yours can too — owned outright, white-labelled under your brand, or run by us. [email protected]