E-commerce / Chargers & electronics
6,900 Products enriched and translated in 7 hours
AITJE built a workflow for a leading Dutch online shop specialising in chargers that supplements product descriptions with technical charging data, translates them for four markets and uploads them directly back into the CMS. Without any external costs per AI token.
Initial processing
- products processed
- 6,900
- first catalogue run
- 7 hours
- GPU-rekenkosten
- $22.33
- translations in the CMS
- 4 markets
The question behind the assignment
Which charger is really the right one for your phone?
Choosing a charger involves more than just the wattage stated on the packaging. Voltage, connectors such as USB-C and compatibility with a specific device determine how quickly you can charge. The same charger may be a good choice for one phone but less suitable for another.
This online shop’s original product descriptions lacked that depth of detail. The information was available elsewhere, but needed to be gathered, combined and translated into something that would genuinely benefit a customer whilst shopping. For 6,900 products, and for several countries.
From product description
The information that was already thereTowards genuine decision support
Which phone charges quickly?
And which ones are less likely?
Technical data is transformed into useful information for the customer.
What AITJE has built
One workflow. From CMS to CMS.
An integrated process for analysis, data loading, enrichment and translation. See what happens at each step.
Step-01
Product data
Via the CMS integration, the workflow retrieves the current products, descriptions and available specifications. This means the analysis begins with the catalogue that the online shop is already using.
This step delivers
Existing texts and product details
Toolcalling, explained in simple terms
AI that can retrieve information.
A tool call is a targeted request from the model to a function within the software. For example, Gemma might ask for charging benchmarks for a specific charger. The software carries out that search and returns the data found to the model.
This gives AI access to up-to-date product information beyond its own model knowledge. It does not need to guess loading performance from an existing description. The software handles the links, data processing, calculations and publication; Gemma analyses the content and writes the explanation.
Three connections in the architecture
The CMS
Retrieve products and existing information, and repost enriched texts and translations.
The GPU VPS
Run Gemma on a separate computing environment, connected to the workflow and the weekly start-up process.
Charging databases
Retrieve external benchmark data on chargers, mobile phones and their relative charging performance.
What the online shop can do now
More than just better copy.
More comprehensive product pages
Existing descriptions have been supplemented with features and loading information that were not previously included in the texts.
A better decision-making tool
Benchmark data feeds into filters and provides insight into charging performance. Customers can see ten devices that charge quickly with a charger, how fast they charge and which models charge more slowly.
Go straight to the online shop
Enriched texts and translations were published directly alongside the products via the CMS link during processing.
Also for customers across the border
In addition to the analysis, a Gemma-based translation workflow was implemented, with Gemma Translator forming part of the chosen approach. This makes the technical information usable for clients in four markets, without having to pay for an external model API for each translation.
- DEGermany
- BEBelgium
- LULuxembourg
- FRFrance
The practical benchmark
6,900 products.
One processing round.
The GPU VPS was running 7 hours were spent analysing, enriching and translating the catalogue. Publication to the CMS took place whilst this processing was underway.
The online shop already had a hosting partner due to the high volume of visitors. AITJE arranged a cost-effective GPU-VPS with that provider for the analysis. This meant the existing hosting relationship was maintained, whilst the AI workflow was given its own computing environment.
GPU costs for this run
7 hours × $3.19
$22.33
- Per product processed
- $0.0032
- Average per hour
- ≈ 986 products
- Average per minute
- ≈ 16.4 products
Calculated from the project figures: 6,900 products, 7 hours and $3.19 per hour. This covers only the GPU computing costs; development, storage and other hosting costs are excluded. The throughput applies to the entire workflow, not to a single model call.
Choosing a model
Why did Gemma 4 31B?
For this project, AITJE chose Gemma 4 31B. Combining product information, using targeted tools and creating clear texts were more important than simply the highest generation speed. The chosen A100 provided the capacity to run the model in 16-bit without compression.
Used in this assignment
Gemma 4 31B
Dense model
Content analysis & enrichmentAn alternative
Gemma 4 26B A4B
Mixture-of-Experts
Useful for rapid processing| Benchmark | 31B | 26B A4B |
|---|---|---|
| 2-benchUsing tools · average across three domains | 76.9% | 68.2% |
| MMMLUMultilingual knowledge-based questions | 88.4% | 86.3% |
| GPQA DiamondComplex knowledge-based questions | 84.3% | 82.3% |
| MRCR v2 · 128KFinding information within a longer context | 66.4% | 44.1% |
The alternative, Gemma 4 26B A4B, is a Mixture-of-Experts model. It contains approximately 26 billion parameters, but activates approximately 4 billion per token. This makes it an interesting choice when speed and computational resource usage are more important considerations. The 31B variant is a dense model and scores higher in the published quality tests shown below. This supports the decision to use it for this content enrichment task.
These are Google’s published results for instruction-tuned models; they do not represent our own comparison based on webshop data. Multilingual knowledge-based questions do not measure translation quality. The practical run measures the entire workflow; no separate tokens-per-second measurement is available.
Google · Gemma4’s model mapWhere open models come close
A minor difference. Considered on a per-task basis.
In some tests, an open Gemma model performs similarly to a recent GPT variant. Below are three examples from the comparison with GPT-6, Luna on Low.
| Benchmark | Gemma 4 26B A4B · Reasoning | GPT-6Luna · Low |
|---|---|---|
| Humanity’s Last ExamChallenging questions from a wide range of specialist fields | 19% | 20% |
| SciCodeProgramming for scientific tasks | 40% | 47% |
| AA-LCR v1.1Reasoning with long context | 66% | 74% |
Selection of the three scores that are closest to one another in the provided Artificial Analysis comparison, rounded as in that comparison. Other tests show greater differences. This compares the 26B A4B variant with GPT-6Luna set to ‘Low’; this online shop workflow ran on Gemma 4 31B. It is not a comparison of product descriptions or translation quality.
Artificial Analysis · model comparisonThe difference in computing costs
Millions of tokens.
One server bill.
This processing takes approximately 68–75 million tokens used, including input and output. Gemma ran on its own GPU VPS. The model tokens were not charged a separate API fee: the GPU bill remained $22.33 for 7 hours.
Why does a single product use input tokens multiple times?
On subsequent runs, the model also reads previous messages and tool results. Below is the estimated progression for each product; the history is taken into account anew with every input.
Retrieving the information
- Input
- ≈ 2,050
- Output
- ≈ 80
System prompt (1,200) + product data (850)
JSON tool call for the database
Enrich & translate
- Input
- ≈ 2,730
- Output
- ≈ 1,000
Previous context (2,130) + search result (600)
Enriched text and translation via CMS-toolcall
Finish
- Input
- ≈ 3,780
- Output
- ≈ 20
Previous context (3,730) + CMS confirmation (50)
Final confirmation: product processed
Per product: ≈ 8,560 input + 1,100 output = 9,660 tokens.
For 6,900 products is that 59,064 million inputs and 7.59 million output: 66,654 million in total. The schedule is an estimate; the stated total project expenditure ranges from 68to75 million.
How much would the same token volume cost via an API?
| Model | Input / 1M | Output / 1M | Input costs | Output costs | Total |
|---|---|---|---|---|---|
| Gemma 4 31B This assignment · own GPU-VPS | $0.00 | $0.00 | $0.00 | $0.00 | $22.33server rental |
| GPT-6Sol | $2.00 | $10.00 | $120.51 – $132.92 | $77.43 – $85.40 | $197.95 – $218.32 |
| GPT-6.1Sol | $2.00 | $10.00 | $120.51 – $132.92 | $77.43 – $85.40 | $197.95 – $218.32 |
| GPT-5.6Terra | $2.00 | $12.00 | $120.51 – $132.92 | $92.92 – $102.48 | $213.43 – $235.40 |
| Claude Opus 5.5 | $4.00 | $20.00 | $241.03 – $265.84 | $154.87 – $170.81 | $395.89 – $436.65 |
| Claude Fable 5 | $10.00 | $50.00 | $602.57 – $664.60 | $387.16 – $427.02 | $989.73 – $1,091.61 |
Calculation example: millions of inputs × input rate + millions of outputs × output rate. For ‘68’ – ‘75’ million, we use the same input/output ratio as in the diagram; the actual distribution has not been measured separately. Excludes caching or batch discounts, additional reasoning tokens, tools, VAT and other infrastructure costs. Other models tokenise and process data differently: these are cost estimates, not actual API runs.
The task determines the solution. For this project, we opted for our own Gemma environment for analysis, enrichment and translation. Quality, integration and control are taken into account, alongside price. The table compares the costs with the selected API models.
And DeepL?
A price per character.
DeepL charges for translation based on source tokens, not input and output tokens. The ‘68’ –75 million model tokens – also includes prompts, tool calls and previous context. You cannot invoice that volume as DeepL translation volume.
- API Growth · per month
- € 29.75
- Included
- 1 million characters
- Per additional million characters
- € 22.00
Dutch monthly package, including VAT, verified on 4 2026. This concerns translation only; product analysis, data uploads and CMS publication require a separate workflow.
DeepL · API ratesCalculate the translation volume
Monthly fee + max(0, volume − characters included) × per-character rate. Add together the source characters from all translation requests. The character volume for this assignment has not been specified, so we are not providing a fictitious project invoice for this.
The workflow continues to operate
New products every week.
The first catalogue has been processed. The weekly synchronisation is active and uses the same supply chain for products that are still missing from the analysis system.
- 01
Start weekly
Automatic synchronisation starts the GPU environment for the new processing cycle.
- 02
Retrieve missing products
New products not yet included in the analysis system are added via the CMS integration.
- 03
Enrich and restore
The workflow retrieves benchmarks, generates the new texts and translations, and inserts them back into the online shop.
AITJE established the connection between the online shop, technical data sources and an open model. The result is a recurring process for product information, decision support and translations, with insight into the computational costs and without an external AI charge per token.
Discover our approachOther case studies
Development agency / Assistant, Coder & model management
Ten models. One team. Control and independence.
AITJE set up a shared AI environment for a development agency, featuring Assistent, Coder and ten models to choose from. A shared token pool of €350 per month replaces nine subscriptions to €200. Model usage and token costs per task can be tracked via Asana.
View the approach
Marketing / customer management, tickets & business knowledge
350 clients. One central workspace.
Following a takeover, a marketing company’s client base grew from 150 to 350. AITJE brought together client data, tickets, projects, invoicing information and call transcripts in a central hub. Staff can search for and chat with their business experts, with access based on their role and AI integrated into an existing Mac mini.
View the approach
Marketing agency / level design & in-house coding agency
25 New game levels every week. With no external token costs.
A marketing agency took over an existing mobile puzzle game but ran into difficulties when creating new levels. AITJE analysed the game, built a reusable level design system and linked AITJE Coder to Qwen on its own server. Now, the agency 25 creates new levels every week from a single assignment in OpenCode.
View the approachMOVING FORWARD TOGETHER WITH AI
Your next step begins with a question.
Got an idea, facing a challenge, or simply curious? Tell us what you want to achieve. AITJE will work with you to find a solution.
