Building AI Tools in Ukraine — A Startup’s First 90 Days

We registered TOV ZELTREX in late 2025. 90 days later, we had 43 AI skills automated, 12 scientific papers published, a 16-person team, and a $220K defense technology proposal on the table. This is the honest account — what worked, what didn’t, and why building AI in Ukraine is both harder and more rewarding than anywhere else.

This isn’t a success story with a bow on top. We’re still pre-revenue on the platform side. Our biggest client engagement is a defense subcontract through a partner. We’ve submitted 21 funding applications and received zero approvals so far. But the technical foundation we’ve built in 90 days would take most startups a year — and we did it from Kyiv, during a war, with a power grid that goes down every few hours in winter.

Why Ukraine for AI

The obvious question: why would you start an AI company in a country at war? Three reasons.

1. Talent Pool Nobody Talks About

Ukraine has one of the deepest engineering talent pools in Europe. Before the full-scale invasion, the IT sector employed 285,000+ people and contributed 4% of GDP. Many of those engineers are still here, and many who went abroad are looking for ways to contribute remotely. The talent is world-class — trained at top universities, battle-tested (literally) in production systems, and motivated by something bigger than stock options.

Our team of 16 includes mechanical engineers, embedded systems developers, AI researchers, 3D printing specialists, and production managers. Try assembling that team in San Francisco for the same cost.

2. Defense Tech Demand

Ukraine is the world’s most active testing ground for defense technology. Drones, electronic warfare, autonomous systems, communication networks — the demand is immediate, the feedback loop is measured in days, and the users are people whose lives depend on the technology working. There is no better product-market fit accelerator than existential necessity.

Programs like Brave1 (defense innovation accelerator) and USF (Ukrainian Startup Fund) are actively funding defense-adjacent tech. The ecosystem is young but growing fast — dozens of new defense tech startups launch every month.

3. Cost Advantage (For Now)

Senior AI engineers in Ukraine command $3,000–$6,000/month. The same talent in the US costs $15,000–$25,000/month. Office space, servers, operational costs — everything is 3–5x cheaper. This won’t last forever as the market catches up, but right now it’s a genuine competitive advantage for startups that need to stretch capital.

What We Built

In 90 days, we built three interconnected systems:

NEXUS Platform

Our AI-native operations platform. A Bridge API with 458 RPC methods across 46 adapters, a desktop application, and two PWA surfaces (tablet and phone). NEXUS is the operational brain — every task, metric, communication, and decision flows through it.

The platform hit v1.0.0 with 1,721 automated tests, 73 QA bugs found and resolved (0 open P0/P1/P2), and a 14-job CI/CD pipeline running across 5 stages.

Night Shift

Our autonomous AI development agent. Night Shift runs 24/7, dispatching tasks from a prioritized backlog, executing them with Claude and Gemini, and grading its own output. In its first 25 days of operation, it completed 300+ tasks — research papers, code implementations, test suites, documentation, and system improvements.

Night Shift isn’t a toy. It has a budget governor (3.5M tokens/week), quality assessment (hybrid heuristic + LLM-as-Judge), an evolution engine that adapts resource allocation, and mentoring feedback loops. When quality dropped 31%, we did a full root cause analysis and deployed 15 fixes — documented publicly.

FEELab Production Lab

The physical side. FEELab is our production facility with four domains: LAB (3D printing, cable assembly, production), DESIGN (mechanical and electrical engineering, PCB design), AI (platform development, integrations), and OPS (operations, HR, sales). 11 core team members plus 4 contractors work across these domains.

The lab produces physical prototypes, assembles electronic modules, and handles production runs for defense clients. The AI platform manages lab operations — scheduling, inventory, quality control, production tracking — through the same NEXUS Bridge API that manages software development.

The Numbers: 90-Day Metrics

Challenges

The honest part. Here’s what makes building in Ukraine genuinely hard:

Power Outages

During winter 2025–2026, scheduled and unscheduled power outages were a daily reality. Our development workflow had to account for machines going offline mid-commit. Solutions: UPS systems for critical workstations, cloud-based CI/CD (GitLab runners on Hetzner), and a development culture that commits early and often. Night Shift running on a German server was actually an advantage here — it kept working when the office went dark.

Export Controls and Compliance

Building defense-adjacent technology means navigating a complex web of export controls, dual-use regulations, and partner compliance requirements. Some AI APIs have usage restrictions for defense applications. Some hardware components require export licenses. Every client engagement starts with a compliance review.

Talent Competition with the EU

Polish and German companies actively recruit Ukrainian engineers, often offering 2–3x local salaries plus relocation. We can’t compete on pure compensation. Instead, we compete on mission (“build technology that matters”), equity (real ownership stakes), and technical challenge (where else can you build an autonomous AI agent that evolves its own task allocation strategy?).

Banking and Financial Infrastructure

International payments are slow. Currency controls exist. Opening accounts for a new company takes weeks. Receiving payments from international clients requires documentation that would make a tax lawyer weep. We manage 7 financial entities across 4 legal structures — each with its own tax regime, reporting requirements, and banking relationships.

Air Raid Sirens

Not a metaphor. Real sirens, real shelters, real lost productivity. On bad days, 2–3 hours of work time disappear. The team has adapted — most of us work from home with basement setups — but it’s a constant tax on focus and energy that no Silicon Valley startup deals with.

What’s Next

The next 90 days have three priorities:

1. Brave1 Defense Grants

We have active applications in the Brave1 defense innovation program for drone protection technology and AI-powered electronic warfare analysis. The pipeline is $2.8M+ across 21 applications. Even a 10% success rate changes our trajectory.

2. NEXUS SaaS Launch

NEXUS v1.0.0 is production-ready. The next step is packaging it for external teams — onboarding flows, multi-tenant architecture, documentation, and pricing. Target: first paying SaaS customer by Q2 2026.

3. Team Scaling

We’re hiring in three domains: AI engineering (Night Shift evolution, Bridge adapters), defense systems (drone protection, EW), and sales (international market development). If you’re a Ukrainian engineer looking for meaningful work, or an international partner looking for defense tech collaboration, reach out.

Building AI in Ukraine isn’t for everyone. But for the right team, with the right mission, there’s no better place to build technology that actually matters.

Lessons Learned

1. Ship fast, fix publicly. We published our quality failures alongside our successes. The 31% quality drop analysis got more engagement than any feature announcement. Transparency builds trust faster than polish.
2. Automate everything you can. 43 AI skills isn’t a vanity metric. Each one eliminates a manual process that would otherwise consume human hours. When your team is 16 people doing the work of 50, automation isn’t optional.
3. Let AI work while you sleep. Night Shift’s 300+ tasks in 25 days represent work that would have taken 2–3 additional engineers. The autonomous agent isn’t replacing humans — it’s multiplying them.
4. Publish research. 12 papers in 90 days established academic credibility that opens doors with funding bodies, defense partners, and international collaborators. Papers are the ultimate “show, don’t tell.”
5. Build for resilience, not convenience. Every architectural decision assumed the worst case: power goes out, internet drops, team member is unavailable, server gets hit. The result is a system that actually handles those scenarios.