Soumit.ma: Reading Every Moroccan Tender and Award Report to Find What the Portal Does Not Show
Why Standard Cloud Stacks and Legacy Tools Failed
The portal publishes thousands of notices a month, mostly purchase orders that close within days. A supplier cannot read them all, and the information that decides whether to bid — who usually wins at this buyer, how many competitors show up, how often the buyer cancels — sits in scanned PDF award reports that nobody aggregates.
Custom Low-Latency Architecture & In-Country Execution
We built an ingestion pipeline that pulls every notice, a document analysis that returns a Go / Go with reservations / No-Go verdict with reasons, the bond and the deadline, and a vision extractor that reads award PVs page by page into structured rows: buyer, lot, every bidder, where each one was eliminated, the amount before and after the commission checked it. Buyer and company names are normalised so the same firm aggregates across spellings.
Key Engineering Deliverables
- Daily ingestion of the public procurement portal: tenders and purchase orders, with sector and region tagging
- Tender-file analysis with a Go / Go with reservations / No-Go verdict, reasons, bond, deadline and requested items
- Vision extraction of award PVs into buyer, lot, bidder, elimination stage and amounts, with a confidence score per row
- Company and buyer name normalisation so a competitor is followed across PVs despite spelling and legal-suffix variants
- Buyer and competitor pages: who wins where, how often a buyer cancels, how many bidders a lot attracts
Measured Production Impact
- At one university hospital, a single supplier holds 135 of 506 attributed lots — a supplier sees this before spending a week on the file
- 28% of lots drew exactly one bid: a second credible offer on those lots is often uncontested
- 88 bids had their amount corrected by the commission for arithmetic errors — a checkable, avoidable way to lose
- Purchase orders close in 3.7 days on average, so the alert has to arrive the day the notice is published
Architectural Context & Deployment Narrative
Soumit is built for any Moroccan supplier that answers public tenders. The first version only read tender notices. The interesting part turned out to be the other end of the process: the award reports. They are published as scanned PDFs, one per tender, and they record every bidder, the stage at which each was eliminated and the amounts before and after the commission checked them. Read one at a time they say little. Read together — 2,885 award notices from 316 buyers, 5,613 bids across 1,139 lots, extracted page by page with an average confidence of 0.97 — they show which buyers rarely conclude a tender, which lots nobody competes for, and where one supplier wins most of the time. Those signals feed the verdict a supplier sees before deciding whether to bid. Figures measured on the production database on 30 October 2026.
Have a similar engineering challenge?
Speak directly with the lead systems architect behind this deployment.
More Production Case Studies
Vext Voice: Morocco’s Sovereign Darija & French AI Voice Telephony Engine
Casablanca bare-metal telephony & WhatsApp VoIP gateway executing Moroccan Darija and French conversational agents with sub-400ms voice latency, direct IAM/Orange/Inwi SIP trunks, mid-call GPS & CMI multimodal actions, and 0.34 MAD/min runtime economics under 100% CNDP Law 09-08 compliance.
AI Email Intelligence: From a Distributor’s Inbox to Tracked Obligations
56,457 emails from 22 mailboxes, going back to February 2017, read by an LLM and turned into structured facts (intent, counterparty, amounts, deadlines, urgency), with 7,939 payment, delivery and regulatory obligations extracted and tracked automatically.