The purchases your procurement stack never sees.
sivra takes the spend you never structured — one-off buys, new-supplier hunts, the messy long tail. An employee asks by phone or chat, and a fleet of agents shops the open market and figures out who needs to approve.
See how it works ↓Sourcing
Shops the open market
A fleet of vision agents finds real options beyond your catalog — marketplaces with no API, regional sellers, live in-session pricing. Not just the suppliers you’ve already onboarded.
Routing
Learns who really approves
Your documented approval matrix is wrong. sivra routes sign-off to how your org actually decides, and adapts every time someone says “not me — talk to X.”
Learning
Gets sharper with every order
Every resolved request is a reward signal. The fleet and the delegation router retrain continually — Pioneer fine-tunes the routing, the Modal-sandboxed agents learn the marketplaces — so each search and each sign-off lands closer than the last.
The wedge
The long tail you can't structure.
Strategic-procurement tools like Tacto and Lio make the spend you've already organized cheaper and cleaner. But you can only optimize what you've structured — and most of what a company buys was never structured at all.
Already structured
What incumbents optimize- Onboarded suppliers & framework agreements
- ERP line items and PO templates
- Negotiated catalogs, known prices
- Recurring, high-value, planned spend
Not yet structured
What sivra handles- Tail spend & one-off, ad-hoc buys
- New-supplier discovery on the open market
- No catalog, no API, live in-session pricing
- Where employees go rogue — maverick & dark spend
When there's no process to route into, people improvise — a personal card, a random vendor, a Slack message that never becomes a record. sivra gives that spend a front door: go to the open market, find the options, figure out who needs to approve.
A swarm of small agents, not one big browser.
sivra dispatches many small, specialized vision agents into open marketplaces in parallel — each cheap enough to spawn on demand, reaching well past the suppliers you've already onboarded.
That's breadth and speed a single large browsing agent can't match: a hundred narrow agents shopping a hundred sites at once beat one generalist clicking through them in sequence.
3
Small — quick look
12
Medium — broad sweep
100
Deep — full market
Up to 100 agents per search — the supervisor sizes the fleet to the request.
Each agent pushes live tiles as it browses; the supervisor watches all of them and collapses the run into one answer. The artifact you get back isn't a list of tabs — it's a decision.
Documented matrix
Learned delegation graph
Now routes
Tools & equipment, €500–5k → Lead Engineer
model v7 · retrained from 41 resolved requests
It learns who actually signs off.
Every company's documented approval matrix is wrong. Who really decides is tribal knowledge — the manager who's technically the approver but always defers to the lead engineer, the budget owner who's been on leave for a month.
sivra starts from sensible priors — org chart, spend limits, category ownership — and adapts from every “not me, talk to X.” Over time it discovers the real delegation graph: the routing your org runs on, not the one in the handbook.
- Priors from org chart, approval limits and category ownership
- Each resolution is feedback — corrected role, corrected urgency, “route to someone else”
- The router is fine-tuned and continually retrained on Pioneer, so the next request routes smarter
Under the hood
How a request becomes a receipt.
No black box. A supervisor agent orchestrates the fleet, aggregates an answer, and hands a clean decision to a router that keeps getting better. Every step writes an append-only audit row.
Intake
An employee asks by phone (ElevenLabs ConvAI) or in the chat UI. sivra creates an org-scoped Order and starts the audit trail.
Supervisor plans
A supervisor agent reads the goal and budget, decides how many agents to spawn and where, and narrates progress to the audit trail as it goes.
The fleet shops
N small, specialized vision models run in parallel on Modal sandboxes, each browsing one marketplace and pushing live tiles to Mission Control.
Aggregate to a report
The supervisor collects each agent's best candidate and synthesizes one comparison report: best option, price-vs-budget, alternatives, recommendation.
Decide
In budget and the requester is authorized? Auto-buy, no human. Otherwise build a DecisionRequest, ask the learned router who signs off, and dispatch.
Resolve & retrain
Approve, counter (re-research with a refined goal), or decline. Every resolution becomes training signal — the router retrains on Pioneer.
Give your tail spend a front door.
Stop the maverick spend by making the right path the easy one. Ask sivra — the fleet shops the open market, the router finds the approver, and every purchase leaves a trail.
