01
RAG with source attribution, not free generation
Every AI answer points at the concrete document and the exact passage. People only trust the AI if they can verify — and in construction that's not negotiable.
Own product Valiro
Valiro: an AI-powered project platform for construction and engineering — from green field to job site.
Valiro connects office and field teams at construction, plant engineering, and engineering firms. A platform I work on daily — and one that flows directly into my consulting for industrial mid-market clients.
- iOS · AndroidMobile platforms
- < 800 msField search latency
- 100%Offline-capable
- EU onlyData hosting
The problem Valiro solves
At construction and engineering firms two worlds run in parallel: the office holds projects, specs, contracts, and documents across an ERP, a DMS, mail inboxes, and shared drives. On the job site, in the workshop, at the customer's location stands the field worker — with no sensible way to access exactly the information they need right now.
The result: phone calls back to the office, half-hour waits, wrong parts, double work, frustrated staff. Existing solutions (Microsoft Teams, SharePoint, dedicated field-service suites) are either too generic, too cumbersome, or too expensive for mid-sized construction and engineering firms.
Valiro picks up exactly there: AI-powered search across all project and document corpora, mobile-first, offline-capable, with clean permission filters. People ask in natural language — the platform returns the right spec, the right photo, the right person to call.
How we built it
Valiro was designed from day one as a mobile-first product. Not 'also on mobile' — but: the central user story is a field worker standing in the workshop, one hand free, needing fast answers. Everything else follows from that constraint.
Architecture: native iOS and Android apps on the field side. A web app for the office. Both connect to a shared API that orchestrates a vector store for AI search, classic Postgres for structured project data, and connectors to common ERPs. RAG with curated source references — so the AI doesn't hallucinate but points at the right PDF.
The hardest part wasn't the AI. The hardest parts were the operational ones: offline sync that resolves conflicts cleanly. Permissions that work end-to-end on document and project level. Adoption by people who've used different tools for twenty years. Performance on older phones in patchy LTE.
Architecture decisions that paid off
02
Offline-first with conflict resolution
The app works without network. Sync is intelligent, with clear conflict rules and a UI that informs the user about conflicts rather than silently overwriting data.
03
ERP connectors instead of ERP replacement
We don't replace existing ERPs. We read from them, write back selectively, and respect the system of record. That has shortened the adoption path with mid-market clients enormously.
04
EU-hosted models for compliance
Models run in the EU, DPAs are clean, data flows documented. For construction/engineering firms with public-sector clients, this is often a precondition.
05
Mobile performance as a KPI
Field search latency is a KPI, not a detail. We measure response times under LTE-patchy conditions, not in fiber-optic offices.
06
Multi-tenant from day one
Even when the first customer was the only one: multi-tenancy is in from day one. Migrating a single-tenant to multi-tenant is one of the most expensive architecture decisions — we avoided it.
Technology stack
Frontend mobile
Swift / SwiftUI Kotlin / Jetpack Offline sync
Frontend web
TypeScript React Next.js
Backend & data
Node.js Postgres + pgvector S3-compatible
AI & models
RAG pipeline EU-hosted LLMs Embeddings
What consulting clients take away from Valiro
01
Adoption beats every model debate
Which LLM you use is usually the least important question. What matters is whether people actually open the solution — and that depends on UX, performance, and permissions, not the model.
02
Source attribution is non-negotiable
In industrial, engineering, and construction settings, 'the AI says' is not acceptable. Every answer needs a verifiable source. That constrains the architecture — and prevents costly mistakes.
03
ERP integration is mandatory, not optional
An AI solution that runs parallel to an existing ERP will die in industrial mid-market firms. It has to read from the ERP and write back selectively — otherwise you create shadow data.
Let's talk for 30 minutes.
I listen, ask questions, and tell you honestly whether and how I can help.