Sovereign Vertical AIIn active development · public previewComing soon · Skylar Cobol 4B

Legacy Coding Agent

Skyl4r builds proprietary vertical AI models. The first — Skylar Cobol — reads, explains, fixes and modernizes mission-critical legacy code without a single line ever leaving your infrastructure. Legacy is the entry point, not the ceiling.

+220B
Lines of COBOL still in production
570+
Open-model downloads in the first month
Live
Real and live — in active development, not a finished product
The making of

How a Skylar is born.

From raw legacy code to a compiler-verified specialist — every stage happens in-house, on our own infrastructure. Nothing borrowed, nothing wrapped.

01
Data
02
Tokenizer
03
Pre-train
04
SFT
05
Ship
Under the hood

The whole stack, built in-house.

From raw data to a compiler-verified answer, every layer under Skylar is ours — and you can inspect it. Click a layer to see what it does.

In-houseLAYER 01/06

Proprietary data pipeline

A curated, clean-licensed COBOL corpus, organized by domain — the fuel behind every training run. No third-party datasets, no scraped code of unknown provenance.

The Product

Skylar Cobol, at work.

A rich-console workspace where the model reads, fixes and verifies mission-critical COBOL — shown here on a classic production failure.

SKYLAR COBOL · SESSION 0847 · ON-PREM
USER> Nightly batch INTFEE01 abends with SOC7 since the August
release. Can you find why?
SKYLAR> Reading INTFEE01.cbl (1,214 lines) — done.
The abend is a data exception in the COMPUTE at line 402:
402 COMPUTE WS-FEE ROUNDED = WS-BAL * WS-RATE / 100
WS-RATE is PIC S9(3)V99 COMP-3, loaded from RATE-FILE.
The August release added rows where RATE is spaces
packed-decimal garbage → SOC7 on first arithmetic use.
Proposed fix (behavior-preserving):
+ IF WS-RATE NOT NUMERIC
+ MOVE ZERO TO WS-RATE
+ PERFORM 9000-LOG-RATE-WARN
+ END-IF
COMPUTE WS-FEE ROUNDED = WS-BAL * WS-RATE / 100
✓ cobc -fsyntax-only: 0 errors · impact: 2 callers checked
MODEL SKYLAR-COBOLCONTEXT [▰▰▰▱▱] 2.3K/8K
Reads the estate
Ingests the real program — 40-year-old dialects, copybooks, packed decimal — and explains the failure in plain language.
Fixes like a senior
A behavior-preserving patch with impact analysis on every caller — not a generic AI guess.
Verifies before you ship
Every answer is checked against a real COBOL compiler before it gets near production.
Already real — not a mockup
The open models are published and downloadable — 570+ downloads in the first month, before any marketing.
What we offer

A senior legacy-code expert, on demand — working where your code lives.

Skylar deploys inside your perimeter and joins the team like a specialist who never retires, never forgets, and has read every line of your codebase.

COBOL Assistance
Answers & guidance, on demand
Senior-level support for the teams that keep systems alive: explain any program, answer any question — and onboard new hires in weeks instead of years.
Logic Analysis
What the code actually does
Extracts the real business rules buried under decades of patches — not what the old documentation says, but what production actually executes.
Maintenance
Faster, safer fixes
Diagnoses defects and proposes fixes with impact analysis up front — so a patch in one module doesn't break a batch job three systems away.
Refactoring
Modernize without rewriting
Restructures aging programs in small, verifiable steps that keep behavior identical — removing the risk of big-bang rewrite projects.
Living Documentation
Knowledge that stops retiring
Generates and maintains documentation straight from the source, turning tribal knowledge into a governed, auditable company asset.
API & Cloud Integration
Open the core, gradually
Exposes proven legacy logic to modern channels — APIs, web, cloud services — so the core evolves instead of being replaced.
The company behind the model

One company, many specialists.

COBOL is the entry wedge — highest pain, no sovereign competition — and the recipe is built to be repeated.

Now · ShippedLive
Skylar Cobol
The COBOL specialist — live at skyl4r.ai, open models published, on-prem ready. The recipe is proven in public.
Next · Same recipe
More legacy languages
RPG (IBM i), PL/I, HLASM, Natural: the same data pipeline, training stack and on-prem delivery — sold into the very same accounts.
Later · New domains
New sovereign verticals
The recipe extends to other regulated domains where generic cloud AI can't legally go. The company is the specialist factory, not one model.
Proprietary data pipeline
Our own models
Sovereign delivery
Built for the long run
Open source · from scratch

Not a wrapper. A model of our own.

Skylar is not another agent sitting on top of someone else's LLM. It is a proprietary model trained 100% from scratch — data pipeline, tokenizer, weights — built in Italy. That is what makes sovereign, on-prem deployment actually possible.

100% from-scratch model
Own weights, own tokenizer, own training runs. No third-party model underneath, no API dependency — the whole stack answers to us, and to you.
Open-source framework
The framework Skylar is built and served with is open source on GitHub — you can inspect how the model is trained, packaged and run.
How it ships

Deployed where your code already runs.

On-premise / air-gapped
Your code never leaves the building. The model runs inside your perimeter — nothing is exported, nothing is sent to a third-party cloud.
Secure SaaS
skyl4r.ai — a hosted console for smaller teams and evaluation, so you can see Skylar on real COBOL before any deployment.
Managed service
Delivered with system-integrator partners, inside the modernization programs your organization already runs.

Built on Skylar's own proprietary AI models, purpose-built for enterprise legacy languages — not a wrapper around a third-party cloud.

Try Skylar on your code.

A 4–6 week pilot, on your premises, on your real programs. Measurable from the first weeks: programs explained, documentation generated, fixes verified.

Sovereign · On-prem · Compiler-verified