Fijomi Systems designs, builds and runs AI agents and automated AI flows for organisations that want real return on investment, not demos. We bring more than 30 years of data architecture experience to modern AI, and choose the most economical model for each job, from rules and small fine-tuned models up to frontier LLMs, so you only pay for intelligence where you need it.
Building AI solutions with
Each stage resolves what it can at the lowest cost and passes only the uncertain cases on.
01 — COST
“Every request goes to the biggest model, and the bill shows it.”
Sending everything to a frontier LLM works in a demo but doesn’t scale economically. We design layered pipelines where cheaper stages (rules, classifiers, small fine-tuned models) handle the bulk of the work, and the expensive models only see the cases that need them.
02 — DATA
“We can’t send customer data to a third-party API.”
Personal data (PII) can’t be shared casually, and GDPR applies. Where data must stay inside your network, we deploy open-source small and large language models internally, so your content never leaves your infrastructure.
03 — PRODUCTION
“The prototype worked. Real life didn’t.”
Real processes are full of exceptions. We work alongside your team to map the whole process, both what goes right and what goes wrong, so the solution handles the core flow and the edge cases. We decide deliberately where automation is safe and where a human stays in the loop.
Autonomous and semi-autonomous agents that carry out real work: monitoring systems daily, spotting issues, raising tickets and taking action, with clear boundaries on what they can do alone.
AI agent development ➤End-to-end pipelines that ingest data from your sources through connectors, then classify, extract and route it, combining rules, small models and LLMs in the most cost-effective order.
AI workflow automation ➤Fine-tuned small language models (SLMs) and open-source LLMs running inside your own environment, for workloads where data privacy, PII or cost rule out public APIs.
Private AI & local LLMs ➤Retrieval-augmented search over your own documents and knowledge bases, giving fast, context-aware answers grounded in your data rather than the model’s guesses.
RAG search ➤Model Context Protocol (MCP) servers that let AI assistants and agents use your product and data directly, with explicit control over what they can reach.
MCP servers ➤We embed with your team, learn your processes first-hand, and build and refine the solution alongside the people who will use it, so it fits the real process rather than the one on the org chart.
Forward deployed engineering ➤We work with your team to map how the process actually runs, including exceptions, failure modes and edge cases, and where the value and the risk sit.
Deterministic rules where they’re enough, small fine-tuned models for high-volume classification, larger models only where reasoning is needed, and a human in the loop wherever confidence is low or the stakes are high.
We build the agents, flows and connectors. Where real data is scarce or sensitive, we start with synthetic data, then validate against real-world data and correct for the gaps between the two.
Daily automated checks catch issues early. Each week we refine rules and models so more cases are resolved at the cheaper early stages, and the system gets cheaper and more accurate over time.
Removing the biggest source of support tickets.
Analysing Runner Technologies’ support tickets showed that 75% were about installing and updating their core product. We designed and built CLEAN_Update, which every Runner customer now uses to download, install and update both the software and its data.
Fraud prevention for student registration.
Ghost student registrations and FAFSA fraud cost US colleges real money. We’re helping Aletheia Systems develop a two-pronged defence: a confidence score on the data each applicant provides, and ID document upload combined with a facial liveness check. We also support Aletheia with marketing optimisation.
Finding invoices and receipts in a flood of documents, cheaply.
Transit Ledger pulls content from many sources through connectors and identifies invoices and receipts, which make up less than 5% of what comes in. Rather than send everything to a large model, we fine-tuned and benchmarked several small language models to pick the most accurate classifier, starting with synthetic data and then recalibrating after measuring how real-world content differed.
Spam detection that gets cheaper every week.
More than a detection system: AI agents run daily checks for issues and raise tickets automatically. Every week we analyse what reached the LLM stage and turn it into better rules, so more spam is caught in the earlier, cheaper stages.
Connecting a core product to AI agents.
We built a Model Context Protocol (MCP) server so that AI assistants and agents can use Runner’s CLEAN_Address product directly.
Getting each call to the agent most likely to convert.
For call-centre technology start-up Insight Solutions, we applied machine learning to call routing, sending each call to the agent most likely to convert.
A microservice that reads a support ticket’s content and rewrites its title to be clear and concise, so teams can triage faster.
A demo agent that plugs into Zoom webinars so attendees can try a product live during the session.
Built to help find student accommodation in the Netherlands: it watches for listing emails, extracts the links, registers for viewings and books a slot automatically.
Fijomi Systems was founded in 2023 by Ronan Cashell, a senior data solution architect based in San Sebastián, Spain.
After graduating in Computer Science from Trinity College Dublin in 1992, Ronan spent three decades designing databases and data architectures. That covers commercial databases (Oracle, Sybase, DB2, Microsoft SQL Server), open-source ones (MySQL, PostgreSQL, SQLite), Linux, and AWS and Google Cloud, plus building software in C, C++, PHP, SQL and PL/SQL. The work spanned banking, automotive, telecommunications and management consulting, along with helping set up several start-ups. More recently Ronan completed MIT’s Data Science and Big Data Analytics: Making Data-Driven Decisions programme.
The idea for Fijomi came in 2022, while Ronan was training developers in Python at Runner Technologies and realised that AI would soon be doing much of that work. Back then, LLMs could handle seconds of work. Today, agents working inside proper scaffolding can handle hours or days of work, sometimes in coordinated swarms. Fijomi has grown with that curve.
Follow Fijomi Systems on LinkedIn ➤“With AI, the only thing preventing you from doing anything is your imagination.”
Tell us about it. We’ll get back to you to arrange a call about where AI can deliver a return, where it can’t, and what an economical, secure solution could look like. No obligation.
Fijomi Systems 