skilmoAI Skills and Context moderation

Pricing

Every AI tool your team uses is either verified or it isn't.

Unverified AI instructions are dark skills. Skilmo finds them, controls them, and — when needed — makes them smarter.

See what it costs
02 — What it costs

Start cheap. Escalate only when you need to.

A bill of materials for everything your AI stack touches. Most teams begin with the first two — fast, fixed-scope. The specialist tier is custom ML for when base precision isn't enough, and you own everything we build.

Start hereFixed scope · fast decision · no friction
01

Gateway Setup

We deploy the sanitisation layer on your infrastructure. Includes one live team training session. You can also do this yourself — we provide the repo and docs.

€500 – €1,500one-time
02

Dark Skills Audit

We inventory everything your AI stack runs — prompts, tools, skill files — and flag what's unverified. Delivered as a structured report.

€1,000 – €3,000one-time
CalibrateTuned to your data, before any model training
03

PII Detection Calibration

Your data has unique formats. We tune the detection layer to your organisation's specific patterns using your own (anonymised) samples.

€2,000 – €5,000project
SpecialistCustom ML · delivered open to you · no lock-in
04

Classifier Fine-Tunecase study ↓

When standard PII detection isn't precise enough, we train a custom model on your domain. Typically 7 training iterations on GPU. Full project, delivered open to you.

€7,000 – €15,000project
05

VLM Document Fine-Tunecase study ↓

When you process scanned documents or images, OCR errors cause missed PII. We fine-tune a vision model to your document types.

€7,000 – €15,000project
03 — Worked examples

Two typical project costs, in full.

Two real specialist engagements, costed line by line — so you can see exactly where the money goes. The pattern is the same both times: a meaningful fine-tune is roughly 7 training iterations of ~15 GPU-hours each — train, validate, repeat — about 100 GPU-hours end to end. We run it on NVIDIA H100 / H200 GPUs inside the secure ETAIS HPC at the University of Tartu, our preferred sovereign computing partner — nothing leaves that environment. Secure EU compute for a full fine-tune costs a few hundred euros. The expertise is the line item that matters: dataset curation, validation, security auditing, integration.

Example 01 · Classifier fine-tune

Custom PII classifier — HR document processing

A financial-services company processes 4,000+ employee documents per month through an LLM assistant. Standard Presidio detection missed 8% of domain-specific identifiers — internal employee IDs, payroll codes. A fine-tuned classifier now catches them automatically.

What the work was
  1. Data collection & anonymisation — 300 labelled samples from client data
  2. Pre-processing pipeline + format normalisation
  3. Process mapping — which document types, which PII categories
  4. ~7 training iterations on NVIDIA H100 — ~12 GPU-hours each (~85 GPU-hours total)
  5. Validation against held-out test set after each iteration
  6. Security audit of model outputs + edge-case stress test
  7. Deployment integration + regression testing
compute · case-1 · classifier

NVIDIA H100 (80GB HBM3) in the secure ETAIS HPC at the University of Tartu.

~7 iterations × ~12 GPU-hours, train + validate each ≈ 85 GPU-hours. Raw compute: ~€250.

# The expertise is what costs — dataset curation, validation,
# auditing, integration. Secure EU compute is a rounding error.

The alternativeAn Azure AI + SageMaker pipeline for equivalent precision: €40,000+ setup, 6–8 month vendor onboarding, and you don't own the model.

What we build belongs to you. The classifier is open to you at source level — you can inspect it, audit it, extend it. No vendor lock-in. Reusable for future implementations.

Business impact

8%→<0.3%

Miss-rate collapsed. At 4,000 documents a month, that's ~320 previously-exposed identifiers eliminated — every month.

Total project cost€9,000 – €13,000
Example 02 · Vision-model fine-tune

Vision model fine-tune — scanned contract processing

A legal-services firm runs contracts through an LLM workflow for clause extraction. Scanned PDFs produced OCR errors that caused the LLM to miss or hallucinate named parties and dates — a liability problem. A fine-tuned VLM now reads their document formats with high precision.

What the work was
  1. Document sample collection — 200 scanned contracts, 5 document types
  2. OCR error analysis + failure-mode mapping
  3. Annotation of ground-truth text regions
  4. Pre-processing: normalisation, resolution standardisation, augmentation
  5. ~7 training iterations on NVIDIA H200 — ~15 GPU-hours each (~105 GPU-hours total)
  6. Precision/recall validation per document type after each iteration
  7. Security audit + adversarial sample testing
  8. Production integration + latency benchmarking
compute · case-2 · vlm

QLoRA on a 7B vision-language model — NVIDIA H200, same ETAIS / University of Tartu infrastructure.

~7 iterations × ~15 GPU-hours, train + validate each ≈ 105 GPU-hours. Raw compute: ~€400.

# Compare: GPT-4V API on 4,000 docs/month ≈ €800–1,200/month, ongoing.
# The fine-tuned model runs on your infrastructure — nothing per call.

Break-evenAgainst ongoing GPT-4V API spend, the project pays for itself in ~14 months — and you hold an owned asset after that.

Business impact

71%→96%

Named-entity recognition accuracy on their document types. Then €0 API cost per document, forever.

Break-even vs. GPT-4V API: ~14 months. Owned asset after that.

Total project cost€10,000 – €15,000

We are IT security auditors first, ML engineers second. We sign off on deployments — not just build them. If something we build introduces risk, we find it before it ships.

Reviewed against GDPR Art. 30 · E-ITS · NIS2

Start with a Dark Skills Audit