AI automation agency that keeps a person on the send button
Boffin Coders is an AI automation agency that takes one repetitive chore off a small team, whether the team is yours or a client’s. The model reads the documents, sorts the inbox or drafts the reply, and a person approves anything that leaves the building. Where the work allows it, documents are read on the device or on your own server, and nothing is sent, paid or deleted without a person. A first working automation takes three to six weeks.
- Runs on your server or on the device
- A person approves anything irreversible
- Answers cite the document they came from
Running a clinic, a trade or a shop and need a website that brings enquiries? That is our local business line.
1. Read the CRM for enquiries with no reply in 7 days
2. Read each thread
3. Draft a follow-up in your tone
Business process automation with AI, three we already run
Before anything we would build for you, the three we run now. Each one keeps a person on the decision, and each one links to something you can open and check.
Reading documents on the device
PDF Toolkit is one of our own apps. It scans a page and reads the text with OCR on the phone itself, so the document never leaves the device. The same approach reads invoices, forms and delivery notes for a business, on a phone or on your own server.
The person holding the phone decides where the result goes.
Drafts a person approves
Our own new business runs through an outreach tool where the model writes each email into an editor and has no way to send it. A rulebook in code checks every draft against nine blocking rules and fourteen warnings, and a queue the model cannot reach does the sending. The write-up quotes every rule.
A person reads, edits and presses send.
Speech to text on your server
IELTS Builder turns spoken exam answers into text with a Whisper model on the platform’s own server, so no recording goes to an AI vendor. It handled 4,584 submissions in a thirty-day window.
A teacher approves every grade before a student sees it.
On-device
PDF Toolkit OCR
Documents never leave the phone
10
Apps on Google Play
Our own, all installable
2017
Working since
Same owners, same team
4.9
Across 20 reviews
Clutch and GoodFirms
Four AI jobs, four pages
Most people arrive here with a chore in mind, not a technology. Pick the sentence that sounds like yours. Each page says what we have built for that job, what it costs and how we would build the parts only your business needs.
It should answer from our documents
Staff ask questions and the answer is somewhere in your manuals, contracts or policies. That is RAG and private LLM deployment: a model that answers from your documents, names the one it used, and runs on your server so the documents stay there.
- Answers name their source
- Runs on your server
- PDFs and Word files
- No vector database to start
It should do the groundwork in our tools
Someone spends hours reading, sorting and drafting inside the inbox or the CRM. That is an AI agent that prepares the work and stops at a locked step for anything that sends, pays or deletes.
- Drafts, never sends
- Works in your own tools
- Every run logged
- A person holds the send
Customers should get answers on our website
Visitors ask the same twenty questions and your team answers them by email. That is a website chatbot built on your own content, which hands over to a person when a question needs one. The assistant on this site is the working example.
- Answers from your pages
- Links the page it used
- Hands over to a person
- Website first
We have calls or recordings to deal with
Meetings, interviews or spoken answers need turning into text. That is voice AI on self-hosted speech to text: the recording is transcribed on your server, not sent to a vendor. A phone agent comes second, and it says it is an AI.
- Transcribed on your server
- Recordings stay with you
- Disclosure by default
- No voice cloning
What is running, and how we build the rest
Which of our AI systems are live, stated plainly, and how we would approach your chore from the first call through the checks to the handover. For everything else we have shipped, sector by sector, see the industries we build for. Agencies that resell AI work to their own clients can use our white-label AI, built under their brand.
One system, live
Speech to text inside IELTS Builder, on the platform’s own server, with a teacher approving every grade. It is an AI system running for a client that you can open and check today, and you can read how it works in our piece on speech to text that never leaves your server.
Three systems we use every day
The outreach tool that drafts our own sales emails. The assistant on this website, which answers from our own pages. PDF Toolkit’s on-device OCR, in a public app anyone can install. Each one is the pattern we would build for you.
Harder work, scoped with you step by step
A phone voice agent on a live line, a model fine-tuned on your data, or forecasting from your own history. Each one starts as a small proof of concept measured on your examples, and a person still approves every action that cannot be undone. Much of this work sits under NDA, so we walk you through the architecture on the call.
What our AI automation agency takes off your team
Six chores we build for, each one built on what the three systems above already do. The model takes the first pass, and a person keeps the decision.
Reading documents
Invoices, delivery notes, application forms and contracts, read and turned into fields in a sheet or your own system. It runs on the device or your server where the documents are sensitive, and anything it is unsure of is flagged for a person.
- Text read with OCR
- Fields into a sheet or system
- Runs on the device or server
- Unsure items flagged
Drafting replies
Enquiry replies, follow-ups and first drafts of quotes, written from your notes and your price list. The draft lands in your inbox or CRM for a person to edit and send. The same pattern suits clinic letters a clinician reviews, covered on our healthcare software page.
- Drafts from your notes
- Lands in your inbox or CRM
- A person edits and sends
- Your tone, your rules
Sorting the inbox
Incoming email and web forms read, labelled and routed to the right person with a one-line summary on top. The routing rules are yours and written down, so you can see why a message went where it did.
- Labelled and routed
- One-line summary
- Rules you can read
- Nothing deleted
Transcribing recordings
Meetings, interviews and spoken answers turned into text by a model on your own server. The recording is stored where you store things, and the text is what the rest of the workflow reads.
- Whisper on your server
- Recording kept by you
- Text into your system
- No vendor upload
Answering from your documents
Staff questions answered from your manuals and policies, with the source named under every answer. When the answer is not in the documents, it says so rather than guessing.
- Source under every answer
- Says when it does not know
- Your documents only
- Staff or customers
Checking work before it goes out
Rules in code that check a draft before a person sends it: a missing detail, a banned phrase, a recipient on the do-not-contact list. A failed check blocks the send and says why.
- Rules in code
- Blocks with a reason
- Suppression lists honoured
- Every check logged
One chore, four steps
The same four steps whether the chore is reading invoices or drafting replies. The approval step goes in before anything touches a live system, not after the first mistake.
Pick one chore
A 20-minute call, then a written scope for one workflow: what goes in, what comes out, where the data runs and who approves. One chore first, because one that works teaches more than five that nearly do.
Deliverables
- One workflow, written down
- Where the data will run
- Who approves what
- A fixed price
Build it on your examples
We build against a sample of your real documents or emails, not a demo set. You see it working on your own examples before it is connected to anything live, and you tell us where it is wrong.
Deliverables
- Tested on your samples
- A list of what it gets wrong
- Rules written in code
- Nothing live yet
Put the person in
The approval step goes in before the automation touches a live system: an inbox, a queue or a screen where a person says yes. The actions it must never take are left out of its reach, not forbidden in a prompt.
Deliverables
- Approval screen or inbox
- Actions it cannot take
- A log of every run
- A fallback when the model fails
Hand it over
It runs in your accounts, the code sits in your repository and the model keys are in your name. We can stay on by the month, or you can hire AI engineers into your own team to run it.
Deliverables
- Code in your repository
- Keys in your name
- A running-cost estimate
- Support if you want it
Published, not quoted on the call
In USD. Hourly and monthly rates move with the stack and the seniority, from a junior on routine work to a senior on complex builds. Nothing is quoted outside them, and if the number does not work for you, you have saved yourself a meeting.
- An automation build starts at
- US$4,000
- Scoped once, with the lines shown.
- Short pieces of work
- US$25-40 an hour
- By seniority and the work.
- A developer by the month
- From US$4,000 a month
- Month to month, a month’s notice either way.
- Agency sprint
- US$2,000 per two-week sprint
- Under your brand, invoiced after you see the work.
Questions owners ask before they book
Answers about cost, time, what the model is allowed to do and where your data goes. If yours is not here, the call is twenty minutes.
It takes one repetitive chore and gives a model the first pass at it. The model reads the invoices, sorts the inbox or drafts the replies, and a person approves anything that leaves the business. For a team of five to fifty, that usually means one workflow, built in three to six weeks, running on your own accounts, with the approval step designed in from the start rather than bolted on.
Not answered here?
Work we’ve delivered, and what we’ve written
Why your RAG chatbot ignores the rules you gave it
Any automation that answers from your documents has a budget for how much of them the model can read at once. This write-up measures it on the assistant on this site, and shows the rule that went missing when the budget ran out.
Read the write-upCheck it yourself
context budget ÷ corpus sizeBring one chore to a 20-minute call
Tell us the task your team repeats every week. We will say whether a model can take the first pass at it, where the data would run and what it would cost, before you spend anything.