Waydev 15B — Waydev's Own SDLC Model

Waydev 15B is Waydev's own AI model for software delivery data. It runs inside the customer's network, so engineering teams and leadership get AI answers without sending code, tickets or people data to an outside AI provider.

Engineering leaders want to ask plain questions such as "Why did cycle time go up last sprint?" and get an answer backed by their own data. General-purpose AI assistants can do this, but only if the data leaves the company. For many of our customers that is not allowed:

• Security teams block source-code metadata and developer activity from going to third-party AI APIs.
• Works Councils and EU privacy rules limit how individual developer performance data is processed and who sees it.
• Regulated industries (finance, healthcare, public sector) need every system that touches their data to run on infrastructure they control.

General-purpose models also don't have a deep, up-to-date view on Waydev's data model or metric definitions. They guess column names, mix up metrics like lead time and cycle time, and answer confidently with made-up numbers. We built Waydev 15B to close both gaps: a model that knows engineering analytics, and that runs where the data already lives.

What it solves

Waydev 15B gives every team the AI features of Waydev with no data leaving their network and no per-token bill to an AI vendor.

Teams that prefer a hosted model can still choose Claude through Waydev, or bring their own Anthropic or OpenAI key. Waydev 15B is the default.

Built on 10+ years of engineering analytics

The training questions come from more than ten years of Waydev answering what engineering leaders actually ask. That experience is what a general-purpose model doesn't have.

Since Waydev started, we have built the metric definitions, the SDLC data model and the reports that engineering managers use every week. We turned that knowledge into the model's curriculum:

• Real questions. The questions customers bring to support, onboarding calls and executive reviews, such as "Which repos have the slowest review turnaround?" and "Are we spending more time on bugs than on features?"
• Our metric definitions. DORA metrics, PR cycle time, review depth, investment allocation, AI coding adoption and more, each defined the way Waydev computes it.
• Our data model. Pull requests, commits, issues, deployments, teams and AI usage, joined the way our own analysts join them.
• Our privacy rules. How to answer under Works Council mode, when to anonymize colleagues, and what never to reveal.
• Our product. How to build dashboards & widgets, write SQL that works on Waydev's schema, and answer from Waydev's documentation.

Each training question is generated from a real Waydev data set, so the model learns from realistic numbers, team structures and edge cases, not toy examples.

How we trained it

We start from a proven open-source base model and fine-tune it on several thousand checked examples of Waydev work. A new version ships only if it beats the current one on our benchmark.

1. Generate questions. Our dataset builder writes questions from live Waydev data across more than a dozen skill areas: names, counts and values over SQL, date ranges, choosing the right metric, Works Council mode, anonymization, product knowledge, PR review and coding, widgets, SQL generation and tricky or misleading prompts.
2. Get expert answers. A frontier teacher model answers each question using the same tools the production assistant has: SQL over the Waydev schema, the metrics engine and the docs.
3. Check every answer. An answer is kept only if it passes automatic checks: numbers match a reference SQL query, the SQL and widgets are valid, no invented columns, no names leaked when they should be hidden, and a rubric grade for the written answer. Failed answers are dropped, not fixed by hand.
4. Train step by step. Each tool call and each final answer becomes its own training example, laid out exactly as the model sees it in production. This teaches the model when to query, which tool to use, and when it has enough to answer. Each version trains on about 1,500 to 1,650 checked conversations, or about 4,400 step examples.
5. Benchmark before release. Every candidate runs a frozen benchmark of over 100 cases, three times each. It is promoted only if it beats the served model and no category drops by more than 10 points.
6. Package. The winning version is compressed to 4-bit precision (a single file of about 9 GB) and served through an OpenAI-compatible API.

Training happens once, at Waydev, before release. Any customer data used in training was aggregated and anonymized first. No names, code or raw records go into the model.

Private by design

Waydev 15B runs on a GPU server inside the customer's network, next to Waydev on-prem. Prompts, query results and answers stay inside that network.

flowchart LR
    subgraph net["Your secure network"]
        teams["Your teams<br/>Ask questions in the Waydev app"] <--> waydev["Waydev on-prem<br/>App, database, access checks"]
        waydev <--> model["Waydev 15B<br/>Your GPU server, model file 9 GB"]
    end
    model -. "never called" .-x vendors["Outside AI vendors<br/>Anthropic, OpenAI"]

• Runs locally. The model is a single file served by an open-source inference server on your hardware, on-prem or in your own private cloud account. It needs no internet access to answer.
• No outside AI calls. With Waydev 15B selected, Waydev never sends prompts or data to Anthropic, OpenAI or any other AI vendor.
• Your access rules apply. The model reads data only through Waydev's tools, which enforce the same account and permission checks as the rest of the product.
• Works Council mode. When it's on, the model answers at team level and refuses to rank or single out individual developers.
• Anonymization. The model can hide colleagues' names in answers, so a manager sees patterns, not people.
• Auditable. Every AI request is logged in Waydev with the model used, the skill applied and whether a customer-owned key was used.

Built for EU rules, Works Councils and HR

Waydev 15B was designed for European companies, where AI over employee data needs a legal basis, a clear purpose and, often, the Works Council's agreement. Running the model inside the customer's network, with privacy rules trained in, makes that agreement easier to reach.

Why it matters in the EU

• GDPR. Developer activity is personal data. Processing it must be limited to a clear purpose, kept to what is needed, and protected against transfer outside the EU without safeguards.
• Works Councils. In Germany and several other EU countries, the Works Council must agree before a company introduces technical systems that can monitor employee performance or behavior. AI analytics over commits and pull requests falls into that category.
• EU AI Act. AI used to monitor or evaluate workers is treated as high-risk, with duties around transparency, human oversight and logging.

How Waydev 15B helps

Working with your Works Council and HR

• Works Council mode and anonymization are account settings, so the configuration can be written into a works agreement and checked at any time.
• The model's behavior is part of training, not a prompt that can be talked around. Asking it to rank individuals under Works Council mode is refused, and our benchmark tests this on every release.
• The audit log gives HR and the Works Council a record of what was asked and which model answered.
• Because everything runs on-prem, your data protection officer can assess the model like any other internal system, with no outside processor to add to the DPIA.

Waydev 15B is a tool that supports compliance; it does not replace your own legal assessment or works agreement.

Capabilities

Waydev 15B powers the AI features across Waydev. It plans its own tool calls, reads the data it needs, and answers with numbers from your account.

The model introduces itself as Waydev SDLC, built by Waydev on an open-source base model.

Quality

9 out of 10 answers right, with half the mistakes of the base model. On our hold-out benchmark, Waydev's training brought 53% fewer wrong answers than the same open-source model before training (the base model got it wrong more than twice as often).

The model keeps improving. We retrain it continuously with new questions, new skills and harder test cases, and every release has to pass the benchmark gate. Waydev 15B is now on version 3.

The benchmark uses a frozen date and a fixed data set, so every version is scored on the same questions with the same correct answers. It covers the same skill areas as training: numbers from SQL, metric choice, dates, Works Council mode, anonymization, identity, product knowledge, widgets, SQL generation and robustness. Each case runs three times. The suite has since grown to 104 cases, and we use it to gate every new release.

Numbers are checked against reference queries, not judged by eye. A right-sounding answer with a wrong number counts as a failure.

Hardware requirements

One NVIDIA GPU with 16 GB of memory runs Waydev 15B. A 48 GB GPU serves more users with longer context.

Also needed on the GPU server:

• Linux with an NVIDIA driver supporting CUDA 12 or newer
• 30 GB of free disk for the model file and the inference server
• Network access from the Waydev server to the model's port; no internet access needed

Both sizes use the same 4-bit model file. Larger GPUs or several GPUs can serve more users. Training is done by Waydev on 80 GB data-center GPUs and is not needed at the customer site.

Limitations and FAQ

Waydev 15B is a specialist. It is strongest on Waydev data and engineering analytics, and weaker than the largest hosted models at open-ended tasks outside that area.

Is it as good as Claude or GPT? On Waydev questions it is trained for the job, at a fraction of the size and with no data leaving your network. For broad general-purpose work, the largest hosted models are still stronger. Customers who want them can switch providers in Settings → AI Providers.

Does our data train the model? No. An on-prem install never sends prompts or data back to Waydev, so nothing from your servers reaches training. Any customer data used in training was aggregated and anonymized first.

Can it run without a GPU? Not at usable speed. Before each answer the model reads its full context (instructions, tool results, conversation), often 10,000 to 30,000 tokens, and one question takes several such steps. A GPU does this in seconds; a CPU alone takes minutes per question and serves one user at a time. We test and support NVIDIA GPUs. Apple Silicon and AMD GPUs can work but aren't certified.

How do updates work? New versions ship as a new model file after passing the benchmark gate. You replace the file and restart the inference server.

Which languages? The standard model is trained and benchmarked in English. Other languages are available as a custom build (see below).

Customized for your company

We can train a private version of Waydev 15B for your company, using the same pipeline and benchmark gate as the standard model. The custom model is delivered as its own file and runs on the same hardware.

How it works

1. Workshop. We collect the questions your leaders ask, your metric definitions and your privacy rules.
2. Training data. We generate and verify examples from your setup. Any data from your environment is aggregated and anonymized first; no names, code or raw records go into the model.
3. Your benchmark. We add your own test cases to our benchmark. The custom model ships only if it passes them without losing quality elsewhere.
4. Delivery. You receive a private model file that only your installation uses, and updates follow each new Waydev release.