Agents, retrieval and the systems around them, for the processes no packaged product fits.
Case study · European Space Agency
SentiBridge is an agentic toolbox we built for the European Space Agency. The answers operators needed were real but buried in thousands of pages of mission documentation, instrument specifications and schemas. They ask in plain language now, and the answer arrives with its source attached.
The agent answers nothing from memory. It reads the question, picks its tools, runs them, and assembles the answer from what came back. Each tool is narrow enough to test on its own.
An answer about a spacecraft is worthless if you cannot check it. Each response resolves to the document and section it came from, and says so when the evidence is not there.
The people who knew where things were written down were the constraint on everyone else's work. The toolbox turns that into something the whole team can query.
What we build
Grouped by sector, though the shapes repeat across all three. Each earns a build only where the volume, the cost per instance and the decision are genuinely there.
Answered on the first ring, checked against the policy, and in the system of record before the caller hangs up. Anything contested goes straight to a person.
Severity and specialism scored from the notification and the history, with the evidence for each factor attached. It ships in shadow mode first.
Schedules, loss runs and covering emails in no fixed format become checked fields, with a confidence per field and a queue for anything below threshold.
A pass over closed files for missed subrogation, duplicate payment and indemnity leakage, producing a ranked list with the evidence. It does not move money.
Every document present, in date and consistent with the others. The output is a checklist naming the discrepancies. It is not an approval.
Narrative, financials, covenant position and comparable precedent, assembled into the memo your committee already expects with every figure traceable to source.
Alerts enriched and grouped by likely typology, with a draft narrative and a suggested disposition. The analyst's decision stays the record.
A regulatory change goes in; the affected policies, procedures and controls come out, mapped and dated. It will not tell you what to do about them.
Deviations from your playbook, the fallback you usually accept, and the clauses you have no position on yet. That last list is the useful one.
Thousands of documents into a cited, ranked issues list with the extract that raised each one. Gaps are surfaced rather than smoothed over.
Precedent and prior positions out of your own matter files, with citations and permissions intact. Often a Company Brain question rather than a build.
Prior engagements, rates and assumptions into a first pass for a partner to correct. The least glamorous use on this page, and one of the fastest to pay for itself.
Agents
Not a chatbot with a better prompt. It reads a request, picks tools, runs them, checks what came back, then answers or hands over. Real capability, and the most over-prescribed architecture in the market.
The power is the toolbox: narrow, individually testable functions. SentiBridge is that shape. Broad clever agents are hard to debug; narrow tools behind a planner are not.
The market has settled on a few composable skills, mixed per use case. Useful because each can be evaluated on its own, and because confirm is where the risk gets contained.
How much latitude the agent gets should be a setting, not a personality. Generous where phrasing matters, hard-coded where policy does, decided per workflow.
Knowing when to stop is a feature. It routes to the right person and hands over what it already established, so the human starts from the middle of the problem.
Lead enrichment and follow-up in sales. First line support across chat, phone and WhatsApp. Onboarding, reconciliation and exceptions in the back office. Specification questions in engineering.
Retrieval
RAG in one line: find the relevant passages, put them in front of the model, make it answer from those rather than from memory. A demo takes an afternoon. Three decisions separate that demo from something a business can rely on.
Split so a clause or a specification table is never cut in half. In document-heavy work this decides more of your accuracy than the choice of model does.
Fifty candidates from hybrid keyword and vector search, re-ranked down to five. Re-ranking is usually the step where a naive pipeline starts working.
If the evidence comes back thin the system searches again instead of answering anyway. The difference shows up in the cases someone will later ask about.
Retrieval built for one use case is fine until the fourth, at which point you maintain four disagreeing versions of the truth. Then it should become a layer everything reads from.
How you know it works
Whichever job you pick, this is the part that decides whether you can defend it. The layer teams skip is almost always the one that fails silently.