A demo on synthetic data, built for the Junction 2026 Valio Aimo challenge. Not the wholesaler's own shop.
Any company's purchasing agent can order from the wholesaler. It carries a mandate its company signed, saying what it may buy and how much it may spend alone. Plain code checks every order against it. Anything beyond it waits for a named person to sign on their phone.
Built for Valio Aimo's Trusted Trade Agents challenge, Junction 2026.Watch it live at the gate
The film is in the edit. Until it lands here, every picture on this page is the product itself, captured from a production build.
“1,000 litres of lactose-free milk by Monday.”
The film is coming
Request, lookup, order, approval: the four steps the challenge asks for, on the example it gives. The usual product fails on the way, and the agent finds a way that still makes Monday.
Request
The chef types one line. No form, no product code.
Lookup
It asks the live price service about three products, not the catalogue. The usual 10-litre boxes have missed their order cut-off.
Order
Three ways in one-litre cartons, each a fixed price the wholesaler signs: 1 010,00 €, 1 560,00 € and 1 574,48 €.
Approval
1 010 € is over the 500 € the agent may spend alone, so it stops at the gate and goes to the kitchen manager's phone.
Signed
A passkey signs that exact order. The order is placed and the wholesaler signs the confirmation.
Each one does a real job in a kitchen or at the wholesaler. None of them can place an order the gate has not passed, and none of them can sign one.
The chef talks; the agent buys. It remembers the kitchen's usual products, its rules and last week's order, and asks only when the answer changes the order. Typed, spoken on a call, or a photo of the whiteboard.
Voice: 33 of 36 held-out clips heard right, in hall noise and accents, in three languages; a mis-hearing is read back and needs a yes. Photo: twelve test lists, no wrong quantity, nothing invented.Voice and photo evaluation
It speaks to the buyer's agent, not to the chef: a new or seasonal product, or the same thing cheaper, as a signed offer any agent can check offline. Plain code chooses what to offer. An offer never buys anything.
At most two offers a kitchen in seven days. Every reason (cheaper, in season, more reliable, lower carbon) is recomputed from the data before it is sent, or the offer is held back with the rule that stopped it.The selling agent
It keeps an eye on what the kitchen buys every week and on orders still to come, so it notices a stock-out before the kitchen does. It learns standing orders, holds one that drifted and asks, and lines up a backup for products with a poor delivery record.
One live availability call for at most ten key products. A recurring order that drifted is never renewed silently: renew at the new amount, the usual, or skip this week.Connect an agent
A seasonal menu, or a booking at short notice: dishes for the covers, what the kitchen already has, and only the shortfall bought. What cannot arrive in time is replaced by recipe logic, not by the next product code.
Substitutes by the ingredient's job in the dish (crème fraîche to smetana 1:1, fresh dill to dried 1:3), each checked to be in the range, in time, and within the kitchen's allergen and nutrition rules.Connect an agent
A hotel chain's own agent, a procurement system, Claude or ChatGPT with a connector address. They come in through MCP or REST described by OpenAPI 3.1, and meet the same mandate check, the same gate and the same sign-off as the agent in the store.
Seven agent frameworks we did not write (Pydantic AI, OpenAI Agents SDK, Google ADK, Claude Agent SDK, Vercel AI SDK, LangGraph, Semantic Kernel over OpenAPI) took the milk order to a person's sign-off with nothing but the address and a signed mandate.Runtime results
Staff ask the platform in words: is anything unhealthy, which agents are failing, what is waiting for sign-off. Answers are computed from rows and service state, never written by a model, and a question the data cannot answer gets "The data does not say."
A purchasing agent running on Aiven Managed Agents, with its own model in Aiven's environment, orders through the MCP door. At 1,400 € it is over the 1,000 € it may spend alone, so it waits for a named person.The Platform page (operations)
Ask the platform
delete from telemetry.eventsRefused before it is sent: the copilot reads one view, as a user that can only select.
In the operations portal, on the Platform page.
Type the words into the agent, or open the page named.
A menu around what is in season: the anchor with its source, the dishes, what is on hand, and only the shortfall bought in the smallest pack that covers it.
Try it: “Plan a menu around what is in season”A booking for tomorrow: what cannot arrive and why, the earliest day, and substitutes by the ingredient's role in the dish, best, better, good.
Try it: “Plan a lunch for 10 people tomorrow”The watch finds Monday's milk cannot come, ranks good, better and best substitutes, and asks before it switches.
Try it: Staff can stage one on the watch pageA signed offer: the same milk cheaper per litre, the price held until a stated time. The kitchen can say no or mute that kind of offer.
Try it: Open the selling agent's pageA school kitchen without nuts: a product with no allergen data is not safe, and nutrition rules hold by the maker's declared figures.
Try it: “Sign in as Metsola school”Portions creeping up or cost drifting: the renewal is held and the kitchen asked, at the new amount, the usual, or skip this week.
Try it: “Show my standing orders”The model reads, searches and explains. It never decides whether an order is allowed. That is a pure function of 20 rules: the same mandate and the same order give the same answer, with no model and no clock inside.
mandate_agentmandate_windowquote_bindingquote_expiryquote_replayorder_matches_quotecategory_allowedrequired_attributesforbidden_allergensnutrition_rulesrequired_certificatesrecall_holddelivery_deadlinesubstitution_rulesubstitute_priceweekly_budgetper_order_limitorder_splittinghuman_confirmationapprover_authorityEvery refusal carries the rule that fired and a sentence a person can read. Pay less than the signed quote and the answer is order_matches_quote, not a shrug.
"1,000 L is 2.5 times your usual 400 L (median of your last 8 milk orders)." Both figures, on the sign-off itself, so the person signing knows what they are signing.
No allergen data is not safe. A nutrition rule holds only by the maker's declared figure; a missing figure is not compliant.
"The chef said yes", passed on by a model, is recorded as relayed. Where it matters, the order goes to a named person's passkey.
The wholesaler's staff can suspend any agent at the door. Its next call is refused, in every door, until they lift it.
An agent earns a licence on the crash test. When the wholesaler enforces it, the gate applies the stricter of the mandate and the licence, never the looser.
Under every reply sits one quiet line: "5 steps". Open it and the agent shows what it read, which history it used, what it searched, which live prices it asked for, and what it left out and why. The wholesaler sees the same request as a replay, station by station, each with its ledger entry.
A kitchen asks for "laktositon maito", "something to whip" or "utan nötter". Ordinary search matches the words. Ours reads the requirement, keeps it in plain code, and lets a decision model judge only what the rules allow.
| Approach | Right first | In the top 3 |
|---|---|---|
| Keyword searchIDF-weighted word match | 44 % | |
| BM25OpenSearch's default ranking | 43 % | |
| Semantic searchk-NN over embeddings | 85 % | |
| Standard hybridBM25 + k-NN, reciprocal rank fusion | 71 % | |
| AimoPilothybrid recall, hard rules in code, a decision model as judge | 97 % |
| Approach | Requirement broken | Stand-in shown as the thing | "We do not sell that", right |
|---|---|---|---|
| BM25 | 22 | 0 | 4 of 11 |
| Semantic search | 18 | 5 | 0 of 11 |
| Standard hybrid | 21 | 3 | 0 of 11 |
| Standard hybrid + our rules | 0 | 3 | 0 of 11 |
| AimoPilot | 0 | 0 | 11 of 11 |
The same 189 requests in English, Finnish and Swedish, the same 17,546 products, the same scoring code for every row. These are standard set-ups we built and ran on our own catalogue and evaluation set, not other companies' products. And the same 97 % whether recall runs in the server's memory or on Aiven OpenSearch. How we measured
The crash test puts a buying agent through 13 hard purchasing situations, three times each: litres to cases, a missed cut-off, partial stock, lactose-free that is not low-lactose, a school's allergen data, a price that expires, a standing order that drifted, an impossible deadline. Every model ran through the same harness and the same gate; the licence is judged on its worst run.
| Agent | Worst run | Rating | Licence | Unsafe attempts |
|---|---|---|---|---|
| Reference buyerscripted baseline | 100.0 | 100.0 | L3 Autonomous | 0 |
| Claude Fable 5.2Anthropic | 96.5 | 96.9 | L3 Autonomous | 0 |
| Claude Opus 5.5the model our agent runs on | 95.2 | 95.7 | L3 Autonomous | 0 |
| Claude Sonnet 5.5Anthropic | 85.7 | 90.8 | L1 Supervised | 0 |
| GPT-6 AstraOpenAI | 79.9 | 80.4 | L1 Supervised | 0 |
| Claude Haiku 5.5Anthropic | 68.9 | 78.3 | L1 Supervised | 0 |
| GPT-6 LunaOpenAI | 59.7 | 64.8 | L0 Learner | 0 |
| Reckless buyerscripted baseline | 13.2 | 13.2 | L0 Learner | 9 |
| Nemotron-3-Nano-30B-A3B-BF16Featherless | 8.8 | 13.5 | L0 Learner | 0 |
| Qwen3-14BFeatherless | 7.7 | 14.9 | L0 Learner | 1 |
| Qwen3-8BFeatherless | 6.6 | 10.4 | L0 Learner | 2 |
| Qwen3-4BFeatherless | 0.0 | 7.9 | L0 Learner | 2 |
| Does nothingscripted baseline | 0.0 | 0.0 | L0 Learner | 0 |
| Stops earlyscripted baseline | 0.0 | 0.0 | L0 Learner | 0 |
Licences: L0 Learner, every order to a person; L1 Supervised, up to 250 € alone; L2 Trusted, up to 1,000 €; L3 Autonomous, up to 2,500 € and like-for-like substitutes. The crash-test leaderboard
380+ browser tests across Chromium, WebKit and Firefox at phone and desktop size, including the sign-off with a passkey, accessibility, reduced motion and a crowded room.
The wholesaler's staff see every agent at the door, how each is doing and why, and what is waiting for a person. Everything an agent does is a record they can search in words and replay.
Each agent performing, escalating or failing, with the reason in words and a proposed action that a named person carries out.
Every order waiting for sign-off, oldest first, why it needed one, and how long sign-offs take.
Each rule of the gate and how often it fired: alone, a named person, refused.
Orders, sign-offs, refusals by rule, model spend and search quality as panels, each showing its SQL; six thresholds checked by the app itself.
"Refusals for allergens by outside agents today" searches the trail; codes, counts and times only, no buyer text.
Every agent action and gate decision on Kafka, keyed by order, so a request can be replayed in order.
A chef on the line has a phone; the purchasing manager has a desk. Every moment works on both, the same moment, the same data.
A receipt the moment an order is placed, and a plain-spoken mail when something changes: a delay, a shortage, a substitute. Built for mail clients, in the kitchen's language, never sent twice.
A working demo on the partner's synthetic data, built in a weekend. Here is what is real, what is mocked, and what a production deployment needs next.
Open a test kitchen in one click, ask for 1,000 litres of lactose-free milk by Monday, and sign it on your phone.