Open-source agentic AI for transaction monitoring. "Just ask Jooby."
Real-time AML and fraud monitoring, run by Jooby with 174 specialist AI agents called Joobriks, with outcomes a regulator can follow.
A 30-minute walkthrough with the developer: your data shape, your typologies, and the case, audit trail and human in the loop Jooby delivers. No sales team, no slides.
01 · Ask Jooby
No menu to learn, no query language, no ticket to raise. Type the question the way you'd ask a colleague, and the Joobriks do the legwork, showing every step.
Works the case queue. Turns alerts into narratives an investigator, or a regulator, can read.
Turns a typology described in plain words into a candidate rule. Saved only once you confirm it.
Lets anyone ask questions of the data and models. No SQL, no chart config.
Reads the running system so incidents are diagnosed in minutes, not hours.
Guides configuration for teams new to the platform. Every change waits for your confirmation.
Joobriks work through Jube's own service layer, with every check the platform applies to itself. If a claim can't be validated, you're told.
Approval and snapshotting came long before the agents. Joobriks inherit both, so nothing they do sits outside the record.
Every tool call, LLM call, query and routing decision, and what it cost, is kept against the conversation. If something needs explaining, it already can be. Live signals and an AI walkthrough of the flow then show where agent boundaries need tuning.
Your permissions are checked before a Joobrik is offered any tool. The platform decides what it sees, not the LLM. If a Joobrik gets carried away and writes its own SQL, it is rewritten to enforce isolation, by tenant and by function. Nothing it runs in R can leave the sandbox.
The knowledge Joobriks graze on is Jube, described from the inside by the person who wrote it. Not just the product as it ships today, but the discipline it grew out of: a textbook as much as a manual.
Ask Jooby is the expression of all that work for the agentic AI era. Nothing is lost; it is synthesised and made easier to reach. Together with the tools and the Joobriks, it is Jube's received wisdom, one question away.
- 01DocsEvery page, as it ships: configuration, concepts, getting started and the agent and tool inventory.
- 02Source codeWhat Jube does, read from what Jube is.
- 03Database schemaEvery table and column, so queries are grounded.
- 04The genesis libraryA practitioner's handbook of method guides: regression, Bayesian networks, clustering, anomaly detection, boosting. Plus R and data-shaping tutorials, and model validation and governance.
- 05AML and transaction modellingHow to think about, classify and build detection rules, with a fraud and AML glossary. Years of bending the craft to this use case.
It stays inside Jube, under AGPLv3, organised into 117,683 delicious bites. Only Jooby can graze on it, because Jube owns the copyright. An LLM wrapper bolted on from outside goes hungry. Agentic AI without a rich Pasture and real tools simply doesn't work. So it comes back to the same thing: "Just ask Jooby."
For years the juice wasn't worth the squeeze. DeepSeek-V4.1-Flash changed that: sparse expert routing and a compressed context cache make long agentic loops affordable, and bearable to wait on.
Jooby needs three AI models Jube doesn't ship: an LLM, an embedding model and a reranker. The choice is made for you: DeepSeek-V4.1-Flash, BGE-M3 and BAAI BGE Reranker v2 M3. It's tested, tuned and supported as a set, and swapping models is not something we encourage (but you can).
All three are open weights. No black-box API sits between your transactions and your answers, no vendor can retire or reprice the model underneath you, and you can run every one inside your own walls. Serve them from any provider that hosts all three, or from several commodity endpoints. Sovereignty demands it? Point the same three at your own endpoints, wherever suits you:
No agent platform to buy, no orchestration cluster to run. Turn Ask Jooby on and the rest is there.
Built on Microsoft's open-source Agent Framework, inside Jube. Tool calls are method calls, not round trips.
Flash models are not a compromise: the latest, like DeepSeek Flash, are excellent at this work at a fraction of the price. Add prompt caching and tight context, and every call is costed in the record, so you can see where the money goes and set the ceiling.
When no tool fits the question, a Joobrik does not give up or guess. It reaches for something more general, and everything it reaches for runs under the same permissions, isolation and record as the rest of Jube.
Against Jube's own Postgres, rewritten to what your role can see.
For statistics and the occasional program, in a hardened sandbox.
A second opinion on current thinking, via Tavily. One switch turns it off.
Officers, filings and status from OpenCorporates, mid-investigation.
A question passes down the tiers, one choice at each, until it reaches the Joobrik that owns the job. That Joobrik alone holds the tools. Three hops at most.
Click a Herd, then a Gaze Gatherer.
02 · No more rule guesswork
Most rule sets grow by anecdote. Thresholds picked by feel, rarely revisited, until nobody can say why a rule is there, or why it's still there. Then the regulator knocks.
Picks one of dozens of model types, from logistic regression to isolation forests, and trains it on your data from scratch.
The same dozens of model types, plus one unfair advantage: topology after topology, until bigger stops being better. It catches the unknown unknowns a fixed threshold never will.
The stronger model takes over; in a dead heat, the simpler one wins. Every bout stays on record, so why a model is there, and why it's still there, always has an answer. A pile of guessed thresholds becomes one score, and one threshold you can defend.
03 · The platform underneath
Jube was real-time transaction monitoring long before Jooby arrived. Every Joobrik stands on it, and all of it is in the open.
IP intelligence, bureau data, complex enrichment, third-party services: write the integration in C# as an Inline Script, and it runs natively in the invocation pipeline, as if it shipped with Jube. The same goes for anything models and rules can't express.
04 · Nowhere to hide
Open source means keeping pace with attackers in public. Every line is out there to be probed.
Every pull request is tested in full, in the open, before it merges. Master is the release: improvements reach you as they land, not in a yearly upgrade.
An embedded OWASP Top 10 WAF drops hostile traffic at the top of the HTTP pipeline.
Thousands of penetration tests on every merge. Not once a year.
OWASP ZAP, Trivy and static analysis. Results published.
Jailbreaks, prompt injection and malicious content refused in every direction, logged for review. Nothing sent off to be checked.
05 · Quickstart
With Docker installed, Jube is a copy and paste away. It builds from source, so give the first start a few minutes.
- 01Change every valueplaceholders, not defaults Make each key at least as long as the example. What each key does →
- 02Open localhost:5001Sign in as Administrator / Administrator change it on first login
- 03Want Ask Jooby?DigitalOcean Serverless Inference is the default provider, but it can be changed to any OpenAI-compatible endpoint. Paste in your inference key and set DockerComposeEnableAskJooby=True.
- 04Ready for production?The single-node stack is for evaluation. Jube Cluster takes the same platform to scale: highly available, quorum-based, with automatic failover, online updates and point-in-time recovery.
06 · Who runs it?
Financial institutions are building centres of excellence and reducing concentration risk on outside partners, as Thoughtworks observes across banking. Jube, open source down to Ask Jooby, makes that practical.
It takes a short integration project and a small cross-functional team to run it, with the developer out of the critical path. You have the code, so you have the continuity. Support is there if you want it, faster and cheaper than going it alone, but never because you need it.
Zero to pre-production.
$7,500 · fixedFlat monthly. No clock, no overage.
from $350/mo