Explainable, with teeth
Six decidable outcomes, not a confidence score. It abstains when the record is silent.
Three product lines, in production at national scale: supply chain traceability, mobile security, field verification.
And AI-native services, built on the same engineering.
All of it built on one rule: trust is not asserted, it is what survives being checked. Every answer cites what it drew on — and says so when the record will not support one.
Six decidable outcomes, not a confidence score. It abstains when the record is silent.
We ship a mobile security product. Most AI firms have no security practice at all.
Telecom, government distribution and field programmes reaching hundreds of thousands.
We design for where models fail — evaluation harnesses, regression suites, versioned answers, refusal over guessing. No AI theatre.
In each case the person who needs the assurance is not the person holding the evidence. Most software closes that gap by asking you to take its word.
We built a platform that does not.
Every claim carries the grade of the evidence behind it — and that grade travels with the claim to whoever reads it next.
Answers are signed. Records are hash-chained.
Where the evidence is silent, the system says so rather than producing something plausible.
Retrieval that cites at span level, abstention when the record will not support an answer, and an evaluation harness that proves it still behaves. The same machinery runs in all three product lines.
Different markets, same platform underneath.
Supply chain traceability
Digital product passports — issued by Proman, the engine Bangladesh exporters already use — plus compliance framework mapping and a provenance-native ERP, built to the CEN standards cited in the Official Journal in July 2026. On the buyer side, supplier evidence that originates at the factory rather than being inferred from shipping records.
Cybersecurity
Android security and privacy built in research collaboration with BUET — malware scanning, an encrypted vault, a family safety network. Free, ad-free and open source, engineered for low-end hardware and intermittent connectivity rather than for review units.
Trust products
Shohay verifies that aid reached the household it was intended for — captured at the point of delivery, offline where the network fails, signed and reconcilable afterwards. Ontor verifies the caregiver who actually arrived: every shift gated by identity verification and a geofence around the patient's residence.
So they get their own doors.
If you buy
Find suppliers who can prove what they claim, with every claim carrying its grade of evidence. Pull passport data into whatever traceability platform you already run. Your customs entry and your supplier evidence, answerable together rather than reconciled by hand under time pressure.
Provena SourceIf you make
Answer a buyer’s compliance questionnaire in an afternoon instead of three weeks, with the certificate attached to each answer. Issue the product passport European brands are already asking for. Run the business on an AI-native ERP that captures the evidence as a by-product rather than as a quarterly scramble.
Provena TrustThe same engineering that runs our products is available as a service — for companies in supply chain, trade and regulated operations who need evidence their own systems can defend. Not generic AI consulting. The specific thing we have already built four times.
Duty and origin-evidence diagnostic. Two weeks, fixed fee, refunded if clean.
One live AI path graded against an auditability standard.
Ingest, screening, graded provenance, signed dossier.
Grounding, signing, refusal behaviour, evaluation harness.
Passport data into the platform you already run.
The canonical record underneath everything else.
What an agent did, and under whose authority.
One record, mapped onto the frameworks you require.
Monitoring, refreshes, evaluation re-runs.
That is what makes services credible from a technology company and incredible from a consultancy. If we cannot point at where we have already built it, it does not go on the page.
See the nine offeringsEvery vendor now says explainable. Very few can tell you what their system does when it does not know.
An answer resolves to one of six states: answered with spans, answered at a weak grade, abstained at retrieval, abstained at grading, refused on entitlement, or failed the post-check. Most systems collapse all six into one confident paragraph.
Where the record does not support an answer, the system says so. A model that returns something plausible over a silent record has not helped you — it has moved your risk somewhere you cannot see it.
A signed, span-grounded answer from a small model beats an unsourced one from a frontier model, and costs an order of magnitude less per token. Most pilots die on the cost of running them at scale, not on quality. We design for the bill.
Retrieval that respects who is asking, and a signed record of what an agent did and under what authority. As systems move from answering to acting, that record is the whole audit.
That is why the AI work looks the way it does. Retrieval that respects entitlement is an access control, not a feature. Prompt injection, exfiltration through a retrieval path and over-broad agent authority are security problems, handled by people who work on security problems.
Adversarial review before launch is our first engineering principle, and it applies to a model endpoint exactly as it applies to a telecom deployment.
Biometric SIM verification across a national telecom network. Fertilizer authenticity and anomaly detection across a national agro-input distribution network. Field delivery verification for a poverty-graduation programme reaching hundreds of thousands of households. Chain of custody for recycled-plastic supply chains. Digital product passports in Bangladesh apparel ahead of the 2028 EU textile regime.
Engineering in San Jose and Dhaka. Roughly sixty-five people.
On why a merchandiser’s personal phone is the garment industry’s weakest link for payment fraud, and how factories can close it.
On SIM-swap and OTP fraud in a phone-based economy, and what carriers, banks and families can each do about it.
Not principles in the abstract. The operating patterns we keep returning to because they are the difference between systems that hold and systems that fail in the field.
Every product ships under adversarial review before launch. We design as if the system will be attacked on day one, because the systems we build for — importer compliance, fertilizer authenticity, donor verification — are attacked on day one. Security is the first design constraint, not the audit at the end.
Designs that fail in the village, on shaky 3G, or in a patient's home are failed designs — even if they looked good on paper. We test in the conditions our users actually work in, before we test in conditions we designed for.
Our engineers know how an NGO field supervisor logs a household visit, how a care agency confirms a caregiver arrived, how telecom signaling actually behaves under load. The platform is general; the people building it understand the verticals.
We stay with clients for years, inheriting the consequences of our own architectural decisions. The discipline of living with what you built is what keeps the next thing you build honest.
Every claim in anything we build carries the grade of the evidence behind it — self-declared, document-backed, cross-verified, or verified by independent audit with a named auditor and date — and that grade travels with the claim to whoever reads it next. It is also what keeps the AI honest: a model may draft an answer, but it cannot raise the grade of the evidence underneath it.
Our thesis on why the global trade order is being rebuilt around evidence rather than cost — and what that asks of the people who make, move and buy goods.
Read the thesis →How do you know? We build the answer.