Reproducible AI production
for your team, on hardware you control
Documented workflows your team can re-run and check: local-AI automation for invoices, inquiries and internal documents, and ComfyUI pipelines with fixed inputs and a runbook. Accuracy is measured in numbers, and your data is never sent to a cloud AI service.
Three services
AI Automation Diagnosis
A diagnosis report that tells you which tasks are worth automating, before you build anything.
Business Automation Build
Automation for invoices, inquiries and internal documents, built on AI that runs on your own PC.
Internal Document AI PoC
An AI that answers from your internal documents, measured on real data before you decide to adopt it.
Prices are base fees in US dollars, excluding tax. Japanese consumption tax applies only to Japan-based clients. Anything beyond scope is handled with paid options or an individual quote. Invoicing and payment details are given with the quote.
How this differs from Discovery
A diagnosis report: which of your tasks should be automated with AI?
Including the call not to automate. An honest assessment.
Thanks for visiting. This is Miragea · Eliora, building reproducible, private AI workflows for teams.
“I want AI to take work off my plate, but I don’t know where to start.”
“I’m not sure it is safe to put internal documents or client information into a cloud AI.”
For small and mid-sized companies and independent professionals, this report decides whether a task is worth automating before anything is built. If a task is better left alone, I will say so.
What the report contains (PDF, about 12–20 pages)
- ① Task inventory (up to 5 tasks)
- ② Priority ranking, scored on five points: impact, repeatability, data readiness, risk, and fit for local AI
- ③ Local AI (runs on a PC inside your company) vs cloud AI: which suits each task, and why
- ④ Rough estimate of benefit and cost, with every assumption stated (for example monthly volume × minutes per task)
- ⑤ For the top task: the workflow after introduction, and a small proof-of-concept (PoC) plan
- ⑥ A 90-day roadmap
How it works (delivered in 7 days)
- After you confirm the request, I send a short hearing sheet
- You describe the workflow, the kinds of documents and the volumes (no real data needed)
- I analyse and write the report
- One round of follow-up questions after delivery
The hearing is written, so everything is on record and easy to share inside your company. If you want a build afterwards, I quote it separately. A diagnosis on its own is perfectly fine.
Why local AI is the starting point
Data is processed on a PC or server you control instead of being sent to a cloud AI service. Confidential information stays out of external AI, and there are no pay-per-use AI fees. Cloud AI can still be more accurate or faster for some tasks, so the diagnosis compares both fairly.
Track record and verification (stated honestly)
- I have been studying and practising AI since 2022 (about 4 years)
- I have no paid delivery record yet
- In-house verification on synthetic data: invoice amount extraction 24/24, internal-document Q&A 13/13, inquiry classification 14/14 (0 critical misclassifications). These are small tests and do not guarantee accuracy on your data
Everything is done online. I do not visit you and I do not connect to your PCs or network.
What to send after you request
- What you want to automate and how it is done today (bullet points are fine)
- The kinds of documents or data (e.g. PDF invoices, email, Excel) and the rough monthly volume
- Roughly how much time each task takes per week
- The PCs, software and AI tools you use now
- Your preferred timeline
Unknown items can be marked “unknown”. We will sort them out in the hearing.
Handling of information
- No real data is needed. If a sample is useful, please redact personal and confidential details or replace them with dummy data
- How files are exchanged is agreed individually after the first conversation
- Information you provide is used only for this service
Scope
- This service covers the diagnosis and the report only. It does not include building a system
- Cost and benefit figures are rough estimates based on the conditions you describe, not a guarantee of results
- No video or phone meetings; the hearing is written
- If more than 5 tasks are to be diagnosed, use the paid option
- Typos and factual errors in the report are corrected within 7 days after delivery. Major changes of direction are quoted separately
Cancellation
Cancellation terms are given with the quote. Once the hearing has started, we will discuss it according to the work done.
What to include in your inquiry
The diagnosis is a fixed fee. For builds, or more than 5 tasks, please ask. Tell me ① your goal ② the tasks ③ the type and volume of data ④ your preferred timing. Please do not include personal or confidential information.
Do I need to send real data?
No. I diagnose from your workflow and the kinds of documents or data you handle. I may ask for one or two dummy or redacted samples, but please do not send personal or confidential information as it is.
What is local AI?
AI that runs on a PC or server you control instead of sending data to a cloud AI service. It is an option when documents or personal data must not leave your premises. Cloud AI can be better for some tasks, so the diagnosis decides case by case.
Can the conclusion be “don’t automate this”?
Yes. For low-volume tasks or tasks where an error is costly, I will recommend leaving them out or using existing tools first. The goal is to avoid wasted investment.
Can I order only the diagnosis?
Yes. The diagnosis stands alone. Any build proposal is optional and you are free to decline.
Is there a video call?
No. The hearing and the explanation are written (email). That keeps a record that is easy to share inside your company, and the report is supplemented by the follow-up Q&A.
Build AI automation for invoices, inquiries and documents
No monthly AI fees. AI that runs on your own PC.
Thanks for visiting. This is Miragea · Eliora, turning repetitive in-house tasks into reproducible AI workflows, one task at a time.
Copying invoice data, sorting inquiry emails, hunting through manuals. Small chores add up to a lot of time every month. This service takes one of them and builds a tool that automates it with local AI (AI that runs on a PC inside your company, without sending data to the cloud).
Automations you can choose (one of three types)
A. Invoice and form field extraction: reads dates, vendors and amounts from PDFs or images and outputs Excel or CSV. Uncertain fields or mismatched totals are flagged “needs review”.
B. Inquiry email classification: assigns a category; complaints, urgent items and similar are flagged “needs review” and shown separately.
C. Internal document Q&A: answers from PDFs, Word files and the like (about 100 files), with sources. If the documents do not say, the system answers “not in the documents”.
What you receive
- The complete automation tool (including configuration files and source)
- An accuracy report on your own samples
- An operating manual
- 30 days of Q&A support after delivery
When you run it on your side, there are no pay-per-use AI (API) fees.
How it works (delivered in 21 days)
- Pre-request consultation: I review the task, data format and volume, and confirm feasibility and the delivery mode
- Receive samples (20–50 items; personal data redacted or dummy)
- Build and test
- Accuracy check on your samples and a report
- Delivery and walkthrough (up to 2 rounds of fixes)
Two delivery modes
- ① I process it: you send the data and I run it on my own PC and deliver the results
- ② Setup delivery: I hand over the tool and manual and you run it yourself. I never connect to your PC or network (required PC specs are shared beforehand)
Why local AI
Client and personal information never has to be typed into an external AI service. Because cloud AI can still be more accurate or faster in some cases, I check on your samples before delivery.
Track record and verification
I have been studying and practising AI since 2022 (about 4 years). I have no paid delivery record yet. In-house verification on synthetic data: A: amount extraction 24/24; C: 13/13 questions correct; B: category 14/14 with 0 critical misclassifications. These are results on a small synthetic set and do not guarantee accuracy on your data.
Please contact me before ordering
I check your sample format and confirm feasibility and the recommended delivery mode. Please tell me:
- The task to automate and its monthly volume
- The data format (PDF, image, email, Excel, etc.)
- Your PC environment (for delivery mode ②)
- Your preferred timeline
Questions alone are welcome.
What you provide
- 20–50 sample items (personal and confidential information redacted or replaced with dummy data)
- An example of the output you expect (what a correct answer looks like)
Handling of information
- How files are exchanged is agreed individually after the first conversation. Please redact personal and confidential details or use dummy data
- Data is used only to process this job and is deleted from my environment after the support period ends
- If personal or highly confidential data is involved, please discuss the scope first
Scope and notes
- The base price covers one of the three types, with validation on samples
- Multiple functions, integration with other systems, and handwriting or unusual layouts need paid options or a separate quote
- AI output can contain errors. For important work, please keep a human final check
- Fixes: up to 2 within the agreed specification. Spec changes are quoted separately
- I cannot take work that handles large volumes of personal data, needs 24/7 operation or uptime guarantees, or requires medical or legal judgement
- Maintenance and improvements after delivery (from $600 per month, excl. tax) are quoted separately
- I do not visit and I do not remotely connect to your PC
What to include in your inquiry
I check your sample format and tell you whether it is feasible and which delivery mode fits. Please tell me ① the task ② monthly volume ③ data format (PDF, image, email, etc.) ④ timeline ⑤ your PC environment (if you run it yourself). Please do not include personal or confidential information.
Is my data sent to an external AI service?
No. It is not sent to any external cloud AI service. In delivery mode ①, I process your data with local AI on my own PC (so you send the data to me). In mode ②, you run it in your own environment.
Are there monthly costs?
When you run it yourself (mode ②), there are no pay-per-use AI (API) fees or server fees. Electricity for your own PC is yours. Only if you want maintenance and improvement after delivery, it starts from $600 per month (excl. tax).
What PC do I need for mode ②?
It depends on the function and the AI model. A PC with a GPU is often preferable; I confirm requirements in the pre-request consultation. If your PC does not meet them, I will propose mode ①.
Can it read handwriting or complex tables?
Accuracy drops with layout. The base price assumes clean PDFs and images. I validate on your samples and tell you beforehand whether it is feasible.
How accurate is it?
The results above were on my own synthetic data; yours may differ. You receive the validation on your samples as numbers in the report.
Build a proof of concept for an AI that answers from your documents
Answers with sources, and “not in the documents” is measured too.
Thanks for visiting. This is Miragea · Eliora.
“We want to search our rules, manuals and past documents with AI, but putting confidential documents into a cloud AI worries us.” For companies in that position, I build a proof of concept (PoC) with local AI, measuring on your own documents whether it really works before you commit to adoption.
What this PoC does
- Organise the target task and documents (e.g. answering questions about internal regulations)
- Create a set of 30–50 evaluation questions together with you
- Build, with local AI (llama.cpp plus an open AI model), an assistant that searches the documents and answers with sources. If the documents do not say, it answers “not in the documents”
- Measure accuracy on your documents (correct and wrong answers, and how it handles questions that should be answered “not in the documents”)
- Deliver an evaluation report (with error tendencies and improvements) and a proposal for production (configuration, cost, operations). Based on the result I tell you whether adoption, passing, or a combination with cloud AI fits
Included in the base price
- One task and up to about 200 documents (PDF, Word, text, etc.)
- A simple chat screen that runs in the browser
- Evaluation report (PDF) and adoption proposal
- Configuration files and a manual
- 30 days of Q&A support after delivery
Delivery takes about 45 days.
If you need more
Additional tasks or documents, or combining with invoice extraction or inquiry classification, are available as paid options. Large builds or production operation are quoted individually after you contact me.
Delivery modes
- ① I build and evaluate on my own PC and deliver the results and environment
- ② I deliver a setup and manual that you can run yourself (PC requirements shared beforehand)
In both cases I do not visit and I do not connect to your PCs or network.
Why local AI
- Your internal documents never have to be entered into an external AI service
- No pay-per-use AI (API) fees
- Cloud-scale AI can still be more accurate or faster. I measure it in the PoC, and if local AI does not fit, the report says so
Track record and verification
I have been studying and practising AI since 2022 (about 4 years). I have no paid delivery record yet. In-house verification on synthetic data: internal-document Q&A 13/13, invoice amount extraction 24/24, inquiry classification 14/14 (0 critical misclassifications). These are small tests and I cannot promise the same accuracy on your documents. That is exactly why the PoC measures it on your documents.
Please contact me before ordering
The PoC varies with the volume, type and sensitivity of the documents. Please tell me:
- Your goal and the task (e.g. answering questions about internal regulations)
- Document types, number of files, total size, and whether they are confidential
- Expected number of users, timing and budget range
- If you want to run it yourself (mode ②): your PC and GPU environment
Handling of information
- How documents are exchanged is agreed individually after the first conversation. Please redact personal and confidential details or use dummy data
- Documents are used only for the evaluation and are deleted from my environment after the support period ends
- For personal or highly confidential documents, I recommend starting the PoC with dummy data or an excerpt
- If you want a non-disclosure agreement (NDA), please ask beforehand
Scope and notes
- This is a proof of concept. I do not guarantee production operation or availability
- AI answers can contain errors, and accuracy figures are valid at the time of validation
- Fixes: up to 2 within the agreed specification. Spec changes or additions are quoted separately
- Maintenance and improvements after the PoC (from $600 per month, excl. tax) are quoted separately
- I cannot take work that handles large volumes of personal data, needs 24/7 operation or uptime guarantees, or requires medical or legal judgement
- I do not visit and I do not remotely connect to your PC
What to include in your inquiry
The PoC cost depends on the content. Please tell me ① the goal and task ② document types, number of files and total size ③ whether they are confidential ④ expected number of users ⑤ timing and budget range. Please do not include personal or confidential information itself.
What is a PoC?
A proof of concept: a small, real trial before full adoption to check effect and accuracy. It lets you decide whether to proceed or pass, based on numbers.
Will it be less accurate than cloud AI?
Generally, local models are smaller than large cloud AI, so accuracy and speed can differ by task. The PoC measures this on your documents. If local AI is not suitable, the report says so.
Is it safe to send confidential documents?
How documents are exchanged is agreed individually after the first conversation. For highly confidential documents I recommend starting with dummy data or an excerpt. Documents are used only for the evaluation.
When does the price go above the base price?
When you add tasks or documents, or combine other functions. These are paid options. Larger scope is quoted after you contact me.
Can I ask for maintenance after the PoC?
Yes. Maintenance and improvements (from $600 per month, excl. tax) are quoted separately.
Reproducible pipeline package (ComfyUI)
For studios, agencies and in-house teams that need the same result on every run, not one-off images. This is the fixed package already listed on the inquiry page.
- One reproducible workflow on a local bf16/fp16 stack
- Input and output specification
- A runbook your team can follow
- One revision included
One-off image generation is out of scope.
- Base price
- $5,000 USD(excl. tax)
- Delivery
- 10 business days
Start with a conversation
Tell me what the work is and what kinds of documents or data it involves, and I will reply with what is feasible and a proposal. Diagnosis-only requests, and questions about whether to skip AI altogether, are welcome.
Open the inquiry form