How much acceleration are your tokens buying?
LLM Engineering
Deep hands-on experience and internal tooling with OpenAI, Anthropic and open-weights models. We develop agentic frameworks and token optimization tooling.
No black boxes. No vendor lock-in.
Your team owns it when we leave.
Our Work
Agentic systems we have shipped
A reusable module of prompts, skills and tooling that guides an AI agent through the brownfield analysis of a legacy system, from requirements gathering to implementation.
Shipped as a git submodule, so every team plugs the same engineering discipline into its own repositories.
We built one of the first fully agentic voice-to-voice pipelines, with real-time guardrails and dynamic scenario generation driven by each customer's preferences.
At its core is a MAG system — memory augmented generation — a 3-layer representation of the user that combines conversation memory with a temporal and a spatial memory.
A proof-of-concept agentic workflow that operates hundreds of Kubernetes clusters with zero human intervention, integrating Claude Code SDK, temporal.io and Terraform.
The agent reads the state from the observability platform, asks the Platform Engineering team for advice over MS Teams when in doubt, and intervenes on its own when it is not.
What We Do
We plan and build agentic workflows for enterprise environments with real constraints such as GDPR, HIPAA, PCI-DSS, and the audit requirements of a public certificate authority. We focus on token consumption vs. actual acceleration, measured as code quality and rework.
Our proprietary toolkits instrument any LLM tool or agentic workflow to collect token usage and analyze agentic sessions.
They show where tokens go to waste: where a deterministic script works better, or where prompts contradict and repeat each other.
We turn those findings into fixes, and your AI spend starts buying delivery instead of retries.
Where do LLMs actually fit in your company? How do you extract value from years of Jira tickets, Confluence pages and legacy repos — without violating a single corporate policy?
We build multi-step agents that plan, call tools and act, with query discipline (your agent reads only from the repositories you decide), guardrails, and human-in-the-loop checkpoints where control matters.
Fully integrated into your company's permissions and IT policies — the agentic tooling your teams want, approved by the IT your teams answer to.
Our Secret Sauce
The daily toolbox we built over years of client work. It's why our engagements start fast and stay measurable.
Analyzes your agentic sessions end to end: where tokens leak, which pipeline steps can be deterministic instead of LLM-driven, and which prompts drift over time. It's how we tell you how much velocity your tokens are buying — with numbers, not vibes.
The enterprise version of our open-source llm-template-engine. Prompts become versioned, testable templates instead of strings scattered across the codebase — so you can run evals on every change and roll back a bad prompt like you roll back a bad deploy.
Let's talk.
Or email directly: tommaso@croccocode.com