Generic answers to specialist questions
A general model doesn't know your product names, document formats or house rules, so someone corrects every answer.
LLM Fine-tuning: Does a general-purpose model keep missing your terms, your formats or your rules?
We fine-tune language models when a task needs it, and we test first whether it does: often better retrieval, a clearer prompt or a different model is enough. Sensitive data can stay on private, open-weight models, and models can run on small devices with no internet.
- ISO/IEC 27001 and ISO 9001 certified
- We stay until your team uses it daily
The problem
Recognize this?
Data that shouldn't go to a public model
Contracts, financials and customer records can't simply be pasted into a public chat tool.
Costs that grow with every call
A large model on every request works in a demo, and gets expensive at volume.
What we build
What we do with models
Fine-tuning, when it pays
We fine-tune open-weight models on your examples when measured accuracy says it helps, and we have run our own fine-tuning trials.
Tested on your examples first
A shared definition of correct, with a set of your real examples, decides between retrieval, prompting, a smaller model or fine-tuning.
Private, open-weight models
Sensitive data can stay on private, open-weight models, served behind an OpenAI-compatible gateway your software can already call.
On-device and offline
Speech recognition and synthetic voice that run on small hardware with no internet, as in the field system we engineer for ATR, where a local language model is being added in 2026.
Benchmarked on cost and quality
We benchmark models on cost and quality for each task before choosing, and show you the expected running cost up front.
No single-vendor lock-in
OpenAI, Anthropic, Google or open-weight models: whichever fits the task.
Proof
Our on-device AI work
Voice AI and messaging that keep working when the network is gone
When a disaster takes out mobile networks, first responders still need to call for help, share what they see and report it upward. For ATR, one of Japan's leading communications research institutes, we engineer X-FACE: a portable box that creates its own network and runs speech AI on the device itself. A local language model, which pulls the disaster facts out of each report offline, is being added in 2026.
Japanese and English speech recognition, a wake word and a synthetic voice run on the box, radio chat is transcribed live, and field reports sync to Japan's national disaster information systems once any link is available.
The fine-tuned language models are ATR's own, developed with its AI partner. Our part is making them run reliably offline on small hardware.
Disclosure: X-FACE is being commercialised as VICINO AI by NEI Globe Networks, Kyoto, where Wisflux's founder is Director and CTO.
Source: ATR–Wisflux contracts, activity reports and correspondence, 2020–2026
- Since 2020
- consecutive annual contracts with ATR
- 2 countries
- of field drills and deployments with responders, in Japan and the Philippines
From our own back office
A private model server for data that should not leave our premises
In our own back office, a private model server runs open-weight language models on our own GPU hardware, behind an OpenAI-compatible gateway, for data that should not leave our premises.
We benchmark models on cost and quality for each task before choosing, and show the expected running cost up front.
Source: Wisflux internal systems, 2025–2026
- ISO/IEC 27001:2022 and ISO 9001:2015, certified by TNV System Certification, valid to October 2028.
- GDPR-ready practices.
- Read-only access to your data first; write access only when it is earned.
- Background checks for engineers, secrets in a vault, anonymized data outside production, and on-premises deployment where required.
- You own everything: code in your repository from the first commit, running in your cloud or on your premises.
How we start
One workflow first. Real data, drafts first.
- 01
Workflow review call
Our founder, Chandraprakash Sharma, walks through one workflow with you and says plainly whether an agent can take it on, and if not, says that too. There is no charge for this call.
- 02
Agent pilot
One workflow, on your real data, drafts first, with success criteria agreed before we start. Your team compares the drafts with its own answers.
- 03
Roll out, one category at a time
We switch on the standard cases first and train the people who will use it. Hard cases stay with your team until it is confident, and every draft is still approved before it goes out.
We stay until it is used: a project ends when your team uses the system daily, not when it goes live. Every engagement is quoted to its scope.
You see the work as it happens: daily progress notes, a demo every sprint, and a work report behind every invoice showing where each hour went.
Other ways to work with us: a fixed-scope build with milestones, a dedicated team of named engineers, or run and support for systems we built or inherited.
Questions
LLM Fine-tuning: common questions
Do we need fine-tuning at all?
Often not. We test retrieval, prompting and model choice against your own examples first, and fine-tune only when the numbers say it helps.
Can a model run without sending our data to a public service?
Yes. Sensitive data can stay on private, open-weight models, and models can run on small devices with no internet at all.
Which AI models do you use?
Whichever fits the task: OpenAI, Anthropic, Google or open-weight models, with no single-vendor lock-in. We benchmark models on cost and quality for each task and show you the expected running cost up front.
What does a model cost to run?
It depends on the task and the volume. We benchmark models on cost and quality before choosing and show you the expected running cost up front. Every engagement is quoted to its scope.
Who sees our data?
We are ISO 27001 certified. We start with read-only access, use read replicas instead of production credentials, keep secrets in a vault, and can keep sensitive data on private, open-weight models.
Are we locked in if it doesn't work out?
No. You own everything from the first commit: the code lives in your repository and runs in your cloud or on your premises. If we part ways, you keep everything.
Start with one task a general model gets wrong.
Tell us which task it keeps getting wrong for your team. Tell us about one workflow. We'll walk through it with you and say plainly whether an agent can take it on — and if it can't, we'll say that too.
Book Discovery Call