LLM systems from prototype to production
After the production readiness review, or after a pilot I built, I implement the agreed production changes in your repository.
Independent AI consultant · Brescia
Have an AI prototype that works in a demo?
I check what it still needs before daily use, and can then build the fixes in your repository.
Only have a supplier's AI quote? I read it and tell you in writing whether to go ahead, fix it first, or stop.
What I do
For technical teams: I check an AI prototype that works in a demo and take it to daily use. For business owners and managers: before you sign a supplier's AI quote, I give you a written second opinion. No quote yet, only a process where you'd like to use AI? A process diagnosis tells you in writing whether it makes sense and what the first step is.
I decide what the AI may do on its own and what a person must check, so errors and costs don't come as a surprise.
Already have an AI project or quote? Get a written second opinion
Fractional CTO (part-time, external)
I can work as a part-time Fractional CTO as a standalone engagement, with or without an LLM prototype: architecture, code and vendor reviews, and a delivery roadmap.
Code handover
I document how to run and maintain an existing codebase, with tests, operating procedures, and access I am authorized to transfer. Handover can be a standalone engagement.
What I have built
I describe the problem and what I built, with links to the code and supporting evidence.
openfatture
#openfatture
My own project, open source
Problem
Italian e-invoices must be built in the FatturaPA format and sent through SDI, the national exchange system. Usually a subscription service does this and also keeps your clients and archive.
What shipped
My own open-source project: it prepares the e-invoice and checks its format before it goes to SDI. The data stays on your machine.
For readers who want the technical detail: two projects with public code and documentation.
emotional-memory
#emotional-memory
My own project, open source
Diagram description: emotional-memory
I encode affective state into a memory store and run a published benchmark; the claim matrix feeds back into encoding. Dataset ingest and the published artifacts stay on the project page — this preview omits them.
Problem
Can affective state recorded at encoding time improve recall for agents that talk to the same person across sessions? And where does it stop helping?
What shipped
On a published benchmark (Zenodo DOI) built to test emotion labels supplied at storage time, the method beat a standard cosine baseline on LLM-judged answer accuracy. Negative results outside that regime are in the repo: I treat this as evidence for that case, not a universal claim.
orka
#orka
My own project, open source
Diagram description: orka
Channel input is routed into a scheduler that runs a bounded workflow and writes a checkpoint. An interrupted run resumes from that checkpoint. Protocol adapters stay on the project page — this preview omits them.
Problem
Requests arrive from Telegram, Discord, Slack, WhatsApp, and HTTP, but without a single durable queue there is no reliable path from channel input to a tracked, reviewable LLM workflow.
Pre-release runtime
It takes requests from several channels into one prioritized queue and, if it stops, resumes from the last saved step. Pre-release, no SLA. (MCP/A2A support: detail on the project page.)
Common questions
- Who do you build software for?
- I work with technical teams and founders taking LLM systems into production. Usually they already have a working prototype and need someone to address operational failures, output quality, and running costs. I also work with owners and managers of small and mid-sized companies who have an AI quote to check or a process they want an agent to handle.
- Can you work with an existing team and stack?
- Yes. I work with your existing stack and constraints. I can address state and failure handling, introduce evaluation tests before release, or measure and limit costs. I hand over tests and documentation so your team can maintain those checks. A public example is the production readiness review I ran on mklang, my own open-source project (in Italian).
- What proof do you have that your systems work?
- Some of my work is open source and verifiable. The emotional-memory project has a Zenodo DOI (https://doi.org/10.5281/zenodo.19972258) with reproducible benchmarks; orka is public on GitHub. The emotional-memory repository also records the negative results outside the regime where the method helps.
- How do I start a project with you?
- Write to info@gianlucamazza.it with the operational problem. If you have a supplier's quote, the first step is the written second opinion. If you already have a working prototype, the first step is the production readiness review. For part-time Fractional CTO work, I first agree with you what the team assessment will cover. For handover, I first agree with you what the repository assessment will cover.
Have a project in mind?
Write to info@gianlucamazza.it with the operational problem and the systems involved. I reply with the next step. Email first; a 20-minute call if useful.
Technical deep dives
System design notes explain architecture patterns. The architecture of this site shows how model proposals pass checks before publication.