AI for real construction problems
KANDO AI stops construction’s most expensive mistakes — wrong quantities, wrong sizing, clashing designs — before they cost money. Every number comes with its source, formula and unit.
The investment caseHow it works
Source: I7-2011 art. 5.2.4, p. 88 · EN 12831-1 · drawing A-204 rev. C · capture 12 Mar, 09:41
The construction problem
AI isn’t what’s missing. The problem is that today’s AI can’t be used where a mistake costs money or safety.
In plain terms
A wrong unit on a drawing travels all the way to site: wrong order, delay, rework. Most errors are born on paper — that’s where they must be stopped.
10.8 %
That’s how many EU construction firms use AI. In Romania only 5.2 % — last in the EU.
Only 10.8 % of EU construction enterprises with at least 10 employees used AI technologies in 2025, against an economy-wide average of 20 %. Romania comes last at 5.2 %. The cause is structural: the average EU construction enterprise has 3 people — no budget and no staff for enterprise software rollouts.
Eurostat, reference year 2025
48 %
Almost half of all rework comes from incomplete data and poor communication — it’s avoidable.
48 % of rework is caused by poor project data and poor communication. 30 % of respondents say more than half of their project data is inaccurate, out of date, inaccessible or entered twice, and professionals spend 35 % of their time — over 14 hours a week — on non-productive activity. Construction labour productivity has grown by 1 % a year over two decades, against 2.8 % across the economy.
FMI & Autodesk 2018 · sector analyses
98 %?
Today’s AI is just as confident when it’s wrong — and never says which 2 % is wrong.
Generic language models are wrong with exactly the same confidence with which they are right. In a law office the consequence is a bad citation. In a bill of quantities it is a financial loss. In a cable-sizing or heat-load calculation it is a safety problem. Commercial vendors claim 98 % accuracy without independent benchmarks — without saying which 2 % is wrong.
2025–2026 literature · our own market review
If a tool won’t tell you which number is wrong, the risk belongs to whoever signs. KANDO AI reverses that.
ProblemTakeoff and estimating take days, by hand, from drawings — and one mistyped unit runs through the whole bid.
KANDO AIAI-assisted takeoff from PDF/DWG/IFC, with server-side recalculation and a check on every line item.
ProblemCable cross-section, voltage drop, heat load: a wrong sizing is a safety risk.
KANDO AIThe number comes from an engine working with the formulas of the standards; the independent verifier blocks a wrong one.
ProblemSeveral amended versions of the same code are in circulation, and it is unclear which one applies.
KANDO AIA knowledge graph that names the version and clause applied, and flags contradictions with both sources.
ProblemClashes between architectural, electrical and HVAC designs surface only on site — as rework.
KANDO AIMore than 200 cross-trade rules on the IFC model, with BCF issues sent to the designer before construction.
ProblemSite managers estimate progress from photos; payment certificates can be disputed.
KANDO AIRobotic or phone capture aligned with the BIM model, with a per-element status and evidence.
ProblemDistributors take days to produce technical quotes while the installer waits.
KANDO AIVerified calculation built into the distributor's own ERP and web shop through an API.
The solution: verifiable AI
Three roles, strictly separated: framing, calculating, verifying. The language model never writes a number.
Where a number can be calculated exactly, the language model does not estimate. If verification fails, the user gets a warning — not a wrong answer.
In plain terms
The language model only understands the question. The engine produces the number with the standards’ formulas, and an independent verifier recalculates it — if it doesn’t match, it is stopped.
In Romanian, Hungarian or English — from the app, API or MCP.
Breaks the task into steps and picks the right tool.
Standards, drawings, BIM and business data — with sources.
120+ calculation families produce the number. The language model, never.
Unit, formula, range, recalculation. Wrong? It stops.
Every step goes into a tamper-evident log: question, sources, calculation, check, answer.
A wrong result never gets out: only a verified number passes the gate.
Separately from the calculation engine it checks the inputs, the formula, the unit, the order of operations and the consistency of the result. The novelty is not the checking but the blocking architecture: a result becomes an answer only after passing the gate, and a verification failure is itself information delivered to the user.
Every value carries its document, page, formula and unit.
Every quantity, price or sizing carries its source document, page, formula and unit in an exportable, auditable basis-of-estimate report — exactly what the person signing off a payment application needs.
When a standard exists in several versions, it says which one it applied — and flags conflicts.
The Romanian electrical and heating standards were both amended in 2023, so different versions of the same text circulate simultaneously. A generic model answers from either one. KANDO AI names the version and the clause applied, and unresolved conflicts are reported as such, with both sources.
What the camera can’t see is never marked as built — it is labelled “not verifiable”.
The distinction between not built and not observable — two situations with identical visual evidence and opposite consequences. An unobserved area is never reported as complete. This property is not publicly documented in any of the progress-monitoring products we reviewed.
In practice
Pick a role and follow, step by step, what happens from the drawing to a document you can sign. Playback is automatic; you can click any step.
TodayTo bid on an office building, walls, openings and cable trays must be measured by hand from dozens of drawing sheets. It takes days, and one mistyped unit runs through the whole estimate.
ResultTarget: half the takeoff time on the same drawings, and an estimate where every line item is traceable.
TodayTrade designs are made in different programs. Whether the duct fits under the beam, or a cable meets the amended code, is often discovered only on site.
ResultThe clash is found on screen, not on the slab — and every issue points to the exact clause of the code.
TodayWeek after week, walk-throughs and photos decide what has been built and what can be invoiced. Payment certificates can be disputed, and the areas nobody saw stay uncertain.
ResultDated, traceable evidence for every element — facts instead of disputes.
TodayAn installer's request for cables and distribution boards waits for days, because someone has to check the sizing and the products have to be looked up in the catalogue by hand.
ResultSame-day technical answers to RFQs, with verified sizing — inside the distributor's own system.
The product
Web and mobile app, plus an API for other software. Every module ships with stage-by-stage verifiable results.
In plain terms
Taking quantities off drawings is the most time-consuming office job. KANDO Takeoff speeds it up — and the server, not the model, recalculates every quantity.
The brain: takes the question, breaks it into steps and routes each to the right tool.
Accepts natural-language or structured requests, decomposes them into an execution graph and routes each step to the right tool. Strict multi-tenancy, hash-chained audit log, versioned REST API, signed webhooks and an MCP server.
Our own language model that speaks the trade in Romanian, Hungarian and English — on our own servers.
An open-weight model adapted to terminology, standards and calculation explanation in Romanian, Hungarian and English, with a low-latency variant for the edge. Served on our own infrastructure, not in a vendor's cloud.
Standards, catalogues and scanned books in one place — searchable, with sources.
Ingests PDFs, DOCX, XLSX, BMEcat/ETIM catalogues and scanned books, including mathematical OCR. Hybrid retrieval and a knowledge graph that reconciles contradictions. Rights labels on every fragment.
120+ calculations and 200+ clash rules — this is where the number is made and checked.
Over 120 calculation families: quantities, waste and cutting optimisation, cost, heat load to EN 12831-1, cable sizing and voltage drop to the national electrical standard. Over 200 rules for cross-discipline conflict detection.
Reads Archicad, Revit, Allplan and Tekla models and extracts quantities.
IFC2x3/IFC4/IFC4.3 import from Archicad, Revit, Allplan and Tekla; native Archicad connector. Robust quantity extraction from imperfect models, IDS 1.0 validation, BCF 3.0 exchange, COBie export.
A robot or a phone walks the site and makes a signed capture.
Mission planning from the IFC model, execution on a humanoid robot (ROS 2) or with a phone (LiDAR/ARKit). Signed capture manifest, on-device anonymisation, offline buffering on site.
Compares design with reality: element by element — built, not built, not visible.
Temporal alignment of capture and BIM plus probabilistic per-element state estimation, with calibrated confidence, visible fraction, evidence and deviation against tolerance. Human validation is mandatory before contractual use.
Dashboard, BIM viewer and voice control — accessible, and offline too.
Project dashboard, in-browser BIM viewer, conversational answers with citations. WCAG 2.2 AA and EN 301 549, noise-robust RO/HU voice interface, offline mode for field reporting.
From drawing to bill of quantities and estimate, every line with its formula — plus RFQs.
PDF/DWG drawings, images and IFC; scale calibration and assisted recognition of walls, openings, rooms and MEP routes. An assembly engine turns measurements into a bill of quantities, each line's formula passing through the verifier, with an RFQ sent to suppliers.
Connects to invoicing, e-Factura, ERP and e-mail — no retyping.
A canonical multi-tenant data model and automatic entity linking. Connectors for invoicing, national e-invoicing, project management, chat, e-mail, ERP over REST/CSV/XLSX, BMEcat/ETIM catalogues and outbound webhooks. Relationship-based permissions propagate into retrieval.
A public benchmark of 2,000+ tasks that anyone can use to check accuracy.
Over 2,000 items across five categories, with contamination-resistant parametric generators and continuous evaluation in CI. A public set under an open licence with a DOI — a dissemination instrument and a credibility instrument at once.
Software architecture
Every component is proven technology. What’s new is the order: an answer can only leave through the verifier.
In plain terms
A cable-sizing question passes four layers: interface → intelligence (model, engine, verifier) → data (standard, project) → integration (back into the company’s systems).
Where users and other software connect: web, mobile, API, MCP.
The web and mobile application, the versioned REST API, signed webhooks and the MCP server through which third-party software agents consume KANDO tools.
Where the answer is made and checked: model, knowledge, engine, vision.
Domain model serving, knowledge with provenance, the deterministic engine with its verifier, and computer vision for site state and drawings.
Separated, protected data for each customer, with an audit log.
Isolated per-tenant databases with vectors and graph in one system, object storage for large files and an append-only audit log.
BIM, robot and business systems — joined by reliable messaging.
BIM/CAD, site capture and business connectors, with idempotent asynchronous messaging between modules.
Customer data stays in the region, on our own servers — not in a foreign cloud.
The adapted model runs on our own infrastructure, in-region. A real requirement for public infrastructure projects, met by none of the international products we reviewed — all of them vendor-cloud services.
A malicious instruction hidden in a document never reaches the model.
Content from documents and customer data is isolated and sanitised before it reaches the context; the product catalogue is kept out of the prompt and accessed through a tool. Already implemented in the internal prototype.
Quantities are recomputed on the server from raw points — the interface isn’t trusted.
Quantities are never taken from the interface: the server recomputes lengths and areas from raw points and the calibration scale, with bounds validation.
Simple questions go to the small model, complex ones to the GPU — predictable cost.
Simple queries go to the edge model, complex ones to the GPU, with a quality guarantee — predictable operating cost as usage grows.
More users? Add a node — no redesign.
Additional nodes and GPUs without re-architecting, thanks to stateless containerised services.
Open-source foundation, no licence fees. IP in all proprietary components stays with the company.
Hardware
Our own infrastructure for data sovereignty and predictable cost. Brands are references only; procurement is competitive.
In plain terms
A robot or a LiDAR phone walks the same route every week and reports, element by element: built, not built, or not visible.
Climbs stairs, opens doors — walks the same route every week.
~180 cm, climbs stairs (riser ≤18 cm), ≥2 h per battery, ≥2,000 TOPS on-board compute, dexterous hands ≥16 DoF with tactile sensors, ROS 2-compatible SDK, wireless E-stop, CE
Role A research instrument for repeatable capture and for operating doors, hatches and switches along the route. After the project it becomes the physical node of the inspection service.
LiDAR + 360° camera: point cloud and images aligned to the BIM model.
360°×≥59° LiDAR, ≥40 m at 10 % reflectivity, ≥200,000 pts/s, IMU, PTP sync, IP67; 360° camera ≥8K photo / 5.7K video; recording computer with 32 GB RAM and 2 TB NVMe
Role Produces the point cloud and imagery aligned to the BIM model. Mounted on the robot or carried by an operator in a backpack — identical sensor payload, so the experiments stay comparable.
GPU rented in an EU data centre to train the models during development.
1× NVIDIA Hopper/Blackwell class, ≥141 GB HBM, 24/7, inside the European Union, with a data-processing agreement
Role Domain-model adaptation (LoRA/QLoRA), computer-vision training and pilot serving for the whole duration of the project.
Short, high-power runs to generate training data.
8× H200/B200 nodes with NVLink, on demand, approximately 6,300 GPU-hours
Role Runs the 671-billion-parameter teacher model to generate distillation data, ablations and benchmark runs. It is not a serving model.
Our own production server, replacing the rented GPU from month 22.
2× professional 96 GB GPUs (≥192 GB VRAM total), 2× server CPU, ≥256 GB RAM, 2× 4 TB NVMe RAID-1, redundant power, rack colocation
Role Serves the domain model and the computer-vision models in production and replaces the rented GPU from month 22 — the condition for predictable cost during the sustainability period.
Connectors for invoicing, e-Factura and ERP — kept apart from the data.
Mac mini M-series, 32 GB, 1 TB, 10 GbE
Role Runs the connectors to customer systems (invoicing, e-invoicing, ERP, e-mail) and service billing, isolated from the database to reduce the attack surface.
Databases and the small model — fast answers from memory.
Mac mini M-series, 64 GB, 2 TB, 10 GbE
Role Vector and graph databases, the edge model and the platform's containerised services. The 64 GB are needed for the edge model plus in-memory databases.
≈40 TB for point clouds and drawings — enough for 24+ months.
Mac mini + 4-bay DAS, 4×20 TB CMR, RAID-6 (≈40 TB usable), Thunderbolt
Role Point clouds, imagery and drawings. A single LiDAR mission produces 5–30 GB; capacity covers more than 24 months of operation.
Firewall, segmented network, UPS — zero-trust protection.
Firewall ≥2 Gbps with IDS/IPS and ≥1 Gbps VPN, managed switch ≥8×10 GbE, 1,500 VA pure sine UPS
Role VLAN segmentation (production / management / backup), encrypted tunnel access and power-failure protection — the precondition for the zero-trust model.
An encrypted third copy in another location — the 3-2-1 rule.
100 TB, perpetual licences, client-side encryption before upload
Role The third copy in the 3-2-1 rule, physically separated from the active data.
No internet on site: a 5G router and mesh Wi-Fi provide it.
Dual-SIM 5G SA/NSA router with VPN, IP67 PoE mesh Wi-Fi 6 access point, ≥500 Wh power station
Role Building sites have no network: this provides the uplink for the robot and the capture kit and resumable synchronisation of large files.
Digitising the Romanian and Hungarian technical library — formulas included.
A3+, 600 dpi, V-cradle, ≤1 s per scan, hot folder; text OCR and mathematical OCR for formulas
Role Digitises the Romanian and Hungarian technical library — the proprietary corpus is the differentiator that cannot be obtained by scraping the internet.
Development and testing on every class of phone, rugged ones included.
6× M-series laptops with ≥48 GB + 12 4K monitors; a LiDAR-equipped iPhone and 3 Android phones (flagship, mid-range, IP68 rugged)
Role Development, local inference for testing and validation of mobile capture across every device class — the scalable capture channel.
Why not cloud only? GDPR and sovereignty, predictable cost, and scaling without redesign.
Product lines
For contractors, designers and distributors. We start with a pilot with written success criteria, then move to an annual contract.
In plain terms
Designer, contractor and distributor work with the same numbers — and each retypes them. KANDO AI gives everyone their own tool from one verified data core.
From drawing to verified estimate and RFQ — paid per project.
AI-assisted takeoff from PDF, DWG and IFC → verified bill of quantities → request for quotation to suppliers. Every line passes through the verifier and keeps its formula in the basis-of-estimate report.
Micro-enterprises · installers · design offices
Design vs. reality on site, with a report for the client.
Plan-vs-built from robot or phone capture aligned to the BIM model; cross-discipline rules; a voice technical assistant; reports with provenance for the client of the works.
General contractors and installers
Design checking: cross-trade clashes found before construction.
Design checking against cross-discipline rules, IFC import from any authoring tool, BCF 3.0 issue exchange, rule packs configured per office.
Design offices and checkers
Verified calculation built into the distributor’s webshop and ERP.
Verified calculation consumed through the API from the distributor's ERP or online shop, matched against their own catalogue and delivered over webhooks.
Distributors · software partners
Robotic site inspection as a service — we own the robot.
Robotic and mobile progress inspection: missions planned from the IFC model, verified reports with evidence, declared limits of observability, digital as-built dossier.
Service — we own and operate the robot
Accredited training that creates internal champions in companies.
Accredited courses with professional associations and on-site training. It trains the users who will ask for the product inside their firm and feeds the commercial funnel.
All sectors · funnel channel
| Segment | End users | What they use |
|---|---|---|
| Construction and building services — primary segment | Technical director, estimator, site manager, works supervisor, owner-engineer of an installation firm | KANDO Construct, KANDO Takeoff, KANDO Inspect |
| Architecture, engineering and technical design | Structural and MEP designer, BIM manager, design checker, technical assistance engineer | KANDO Engineer, KANDO Takeoff, KANDO Inspect for the as-built dossier |
| Distribution of materials and equipment | Technical estimator, sales agent; indirectly the installers who are the distributor's customers | KANDO Quote through the API, embedded in the ERP and the online shop |
| Horizontal — software vendors and integrators | Technology partner and their own customers | KANDO API and OEM licence, tools exposed through the MCP protocol |
Market and segments
Contractors and installers, design offices, distributors. Through distributors we reach thousands of installers.
In plain terms
Romania ranks last in AI use. That is an open market: companies have no tool that works in their language, with their standards, verifiably.
≈12,600
construction firms with 10+ staff in Romania — plus ~20,000 micro-firms.
firms with at least 10 employees in Romania, plus roughly 20,000 micro-enterprises active in tendering, who enter through per-project use.
≈4,350
design offices with at least three designers — for design checking.
architecture, engineering and technical design offices with at least three designers — the users of design checking and the as-built dossier.
≈870
distributors with a quotation desk — a gateway to thousands of installers.
materials and equipment distributors with a quotation department, through which the product reaches thousands of installation firms indirectly.
What users gain: half the takeoff time, faster design checks, same-day quotes — and a report an engineer is willing to sign.
Execution plan
Every step delivers a concrete, verifiable result. Commercial launch comes in month 24.
In plain terms
Not one big promise at the end, but six verifiable steps: prototype, experiment report, pilot — trackable from the outside.
What and how we measure: specifications, research protocol, threat model.
Functional and non-functional specifications for every module, the research protocol with measured baselines, the data-rights register, the conceptual architecture and the threat model.
Working in the lab: robot, engine + verifier, first language model.
The robot accepted with a safety test and in research use; orchestrator with verified execution, calculation engine plus verifier, the first domain-model version, the knowledge-base prototype and corpus v1.
Everything runs together in a controlled environment — first experiment reports.
Integrated prototype v0.8, robotic capture in a controlled environment, plan-vs-built v0.5, Archicad connector and BCF workflow, first experiment reports.
Three real pilots with robotic site visits, beta app.
Laboratory and field experiment reports, three validation pilots with robotic visits, the application in beta, Takeoff & BoQ v0.8.
Production infrastructure, own GPU server, security testing and public benchmark.
Production infrastructure and our own GPU server commissioned, hardening and internal penetration test, the public benchmark and the open API specification, multi-site validation reports.
Launch: 5 paid pilots, 6 signed contracts, independent audit.
The robotic inspection service in operation, operational validation in three contexts, five paid pilots, six signed commercial contracts, an external penetration test and an independent technical audit.
Direct sales in three regions, first API partner.
Three regions covered; partnerships with professional associations for accredited courses; the first API partner; the first sites on a robotic inspection subscription.
National coverage and entry into Hungary with a Hungarian interface.
Hungarian interface and corpus, local partners, resellers and integrators; distributors offer assisted takeoff to their installers through the API.
More licences at existing customers, OEM integration into partners’ software.
Additional design licences and sites inside existing accounts, as-built dossiers delivered at volume, verified calculation embedded in software partners' own products.
The team
Six full-time, three sector engineers with 15–21 years of experience. Identities are available during due diligence.
In plain terms
Construction AI often fails because developers don’t know the site. Here, the sector engineers write the rules and lead the pilots.
Owns the platform architecture, the API and BIM integration.
Platform architecture, the orchestrator and the API/MCP surface, BIM integration, the unified data model; classification of industrial-research and experimental-development work, R&D quality, hybrid inference routing.
Research lead for the language model, calculation engine and verifier.
The domain language model, the deterministic calculation engine and the verifier, capture-to-BIM alignment and state estimation, drawing understanding, cross-discipline checking, robotic manipulation policies, the research protocol.
Data ingestion, search and computer vision for drawings and sites.
Data ingestion, retrieval and the knowledge base, computer-vision implementation for site state and drawings, entity linking across sources, the evaluation harness.
The takeoff canvas, development infrastructure and pilot coordination.
The takeoff canvas and the BoQ-to-RFQ flow, the web front end, business connectors, the R&D platform and DevOps (environments, CI/CD, GPU job orchestration), benchmark operations, pilot coordination.
Web and mobile app, accessibility and voice control.
The web and mobile application, WCAG 2.2 AA accessibility and the voice interface, the mobile capture app, test automation and research DevOps.
Owns robotic capture and the inspection service.
Robotic capture (ROS 2, vendor SDK, mission planning), execution of the capture and manipulation experiments, application requirements, operation of the inspection service line.
Construction rules, takeoff validation, leads the general-contractor pilot.
Construction rules for cross-discipline checking, validation of AI-assisted takeoff, leading the pilot at a general contractor. Part-time.
Electrical installation rules, leads the distribution pilot.
Electrical installation rules, the catalogue and quotation flow at the distributor, leading the pilot in the distribution sector. Part-time.
HVAC rules, BIM/IFC, leads the design-office pilot.
HVAC rules, BIM/IFC integration, design-office workflows and leading the pilot at the partner office. Part-time.
The Q&A set annotated by our sector engineers doesn’t exist online — it is the hardest asset to copy.
Why it is hard to copy
These barriers are built with time and access, not money.
In plain terms
Anyone can plug in a generic AI. A proprietary technical corpus, engineer annotations, a blocking architecture and real-site robotic experience can’t be bought.
Romanian and Hungarian technical library plus engineer annotations — not online anywhere.
The Romanian and Hungarian technical library digitised with a professional scanner and mathematical OCR, plus the question-and-answer set annotated by chartered engineers with 15–21 years of experience. An international competitor cannot collect it.
The verifier is the system’s foundation from day one — it can’t be bolted onto a chatbot.
An independent verifier that refuses to deliver a non-conforming result is an architectural decision taken from the outset. It cannot be bolted onto a system in which the language model produces the number.
Standard versions and their relationships in a graph — extendable to any country.
Clauses as typed nodes with amends, repeals, derogates and applies-to relations, plus temporal validity. Applicable to any language and any body of standards — and therefore to the next markets.
Stairs, unfinished floors, doors — wherever people walk, the robot goes too.
Stairs, unfinished floors, doors and hatches — the same routes people take. Existing solutions capture reality with rovers or scanners at fixed waypoints, but do not couple it with a layer of verified engineering knowledge.
We publish the category’s benchmark — and lead it.
Over 2,000 items with contamination-resistant parametric generators, released as a public set under an open licence with a DOI. Publishing it creates the standard by which the category is measured — and we lead it.
Distributors and software vendors resell it through their own systems.
An open API with provenance plus OEM licensing: distributors and software vendors resell the capability through their own ERPs and online shops, which lifts the ceiling imposed by a small sales team.
Why this project
We don’t ask for trust; we give criteria: 10 hypotheses with measurable targets, 6 milestones, a public benchmark.
In plain terms
Risk falls step by step: first the lab proves it, then real pilots, then paying customers — every step with a measurable target.
We write down what could fail and the plan B — no promises without a benchmark.
Ten research hypotheses, each with what can fail technically, what happens if the target is missed and what the mitigation is. We do not promise accuracy without an independent benchmark.
Our internal prototype already works: recomputation, auditable formulas, protection.
An internal prototype, built with our own resources before submission, already implements server-side deterministic recomputation, assembly recipes with auditable formula strings, a tool-calling agent and prompt-injection defence.
All proprietary components belong to the company, on an open-source base.
The orchestrator, the calculation engine, the verifier, the model adapters, the corpus and the datasets belong to the beneficiary; the open-source base is under permissive licences.
Classified under the EU AI Act; customer data stays in the EEA.
Explicit classification against Regulation (EU) 2024/1689 on artificial intelligence, a data-rights register for every source, and customer data that never leaves the European Economic Area.
WCAG 2.2 AA and voice control in safety gloves — what matters on site.
WCAG 2.2 AA and EN 301 549, full keyboard navigation, a voice interface in Romanian and Hungarian usable with occupied hands and protective equipment — the real condition of use on site.
Public benchmark, open API, papers — negative results included.
A public benchmark under an open licence with a DOI, an open API specification, conference papers and an open-access article — including negative results.
The full technical dossier (architecture, research protocol, specifications, implementation plan) is available to investors and partners under a non-disclosure agreement.
Request the technical fileContact
We reply within two working days. For investors we arrange a session with the technical team and a demonstration on real data.
Tetarom Ipari Park, 47C1, etaj 1, birou N6, Cluj-NapocaOpen the map