DataDriven: Reading Every Public Tender in Romania, Alone

I spent twelve years as a developer at Microsoft, two of them in the US on the Bing team. In 2021 I burned out and quit. In 2023, with zero experience as an entrepreneur, I founded a company and launched DataDriven, a platform that monitors public tenders across Romania. Today it crawls more than 10,000 sources daily, reads the actual tender documentation with an OCR pipeline I run on a gaming GPU in my office, and serves over 300 paying firms at around 27,000 lei in monthly recurring revenue. I still work alone, and that is not an accident.

An uncle lost in SEAP

The idea came from watching my uncle, a real estate appraiser who worked on public procurement contracts. He would spend hours inside SEAP, Romania’s electronic public procurement system, because the interface gave him six different menus where a tender relevant to his business might be classified. For him there was one primary category where his work obviously belonged, but the authorities filing the tenders did not always agree. So he searched all six, every day.

It got worse after the search. To download a tender’s documentation you receive a .p7s file, a cryptographically signed wrapper that Windows does not open. The standard move, for him and for everyone else in the industry, was to google “p7s open” and land on some piece of trialware. This is the daily reality of the people who sell to the Romanian state.

I did the obvious engineer thing: I wrote a script that watched for new tenders in his domain, crawled the data in Azure, and emailed him whatever appeared. That script was the seed of the company.

A company of one, on purpose

When I decided to build a business, I decided to build it alone. After the burnout I wanted to eliminate human contact from my working life, bluntly put. No cofounders, no meetings about meetings, nobody to negotiate the roadmap with. I wanted to build the application exactly the way I saw fit.

Launching into silence

I founded the company in 2023 and got into a startup program with Azure cloud credits worth up to $150,000. I burned through about $100,000 of it on the AI scripts that crawled and processed everything. In February 2023 I launched, put up some ads, and waited for the server to crash under the load of eager customers.

The server was fine. Nobody converted. The first paying client arrived in April 2023, two months after launch, at 19 lei per month, roughly four euros. For a long time the business was 40 clients at 19 lei each, which is not a business, it is a hobby with invoices. Until 2025 I barely looked at marketing at all; I was heads-down on the crawler, the data pipeline, and the frontend.

The customers taught me the product

The shape of the real product came from two kinds of customers.

The first was a security services firm. Their tender-finding process was a browser bookmark folder organized by county, and every day someone opened the town hall websites one by one to check whether a new tender had appeared. That was a salaried human being clicking through bookmarks daily. They became my first serious contract at 2,000 lei per month for complete coverage of the sites they cared about.

The second group was catering firms, the companies behind school meal programs like the old “cornul și laptele” (now “masă sănătoasă”), where town halls receive budgets for sandwiches and related products. Catering is logistically bound to a county: you cannot deliver school meals two counties over. So each firm needs to watch every town hall in its county, and only those. That constraint made the product’s value concrete: complete local coverage, zero noise from the rest of the country.

Read the documents, not the labels

The deeper I got into tender data, the clearer one thing became: you cannot trust what the authorities write on the tin. The official CPV classification system has around 10,000 subcategories, and misclassification is routine. I have seen food accessories filed under laptops. Titles and descriptions are frequently vague, misleading, or simply wrong.

So DataDriven reads the actual documentation, the PDFs and scans behind each tender, and classifies from the content. I built my own taxonomy of about 180 categories, deliberately down-to-earth, meant to mirror how Romanian companies actually describe their work rather than how a European classification committee does. That is what lets the platform recommend exactly the products requested in a caiet de sarcini instead of shrugging out something generic like “medical equipment.”

The marketing that finally worked

Google Ads was my first channel, and it mostly bought me branding. Conversion was terrible for a structural reason: people searched for construction work, clicked the ad, and landed on a platform full of tenders they were not eligible for. Public procurement has revenue thresholds, employee counts, and certification requirements, and an ad click cannot check any of them.

Cold email became the channel that actually worked, and the raw material was public data. SEAP exposes who has participated in tenders, who has won them, phone numbers, and activity domains. That is a pre-qualified lead list: these firms already sell to the state, they just find tenders badly. The market is roughly 20,000 firms, so I use Instantly.ai at about $150 per month, plus Google Workspace accounts at around $5 per address, to run multiple sending addresses across several domains with automated warm-up, keeping deliverability healthy at a deliberately limited throughput.

Escaping the cloud bill

When the Azure sponsorship expired, reality arrived: 2,500 to 3,000 EUR per month across Azure Search, Cosmos DB, OpenAI, and storage. For a business then doing about 1,500 EUR in MRR, that is not a cost structure, it is a countdown.

The migration happened in pieces. Azure Search, at about $500 per 16GB instance and two instances for availability, became self-hosted Elasticsearch. The web backend moved to Hetzner. Storage moved to Hetzner’s S3-compatible object storage, after I discovered what egress and ingress really cost: my pipeline downloads documentation from SEAP, OCRs it, re-uploads it to storage, and serves it to processing jobs and users, and all that shuttling produced surprises like a 700 EUR ingress line item in a single month. For overflow, when my local infrastructure has downtime, I spin up Azure Batch Low Priority instances for a few hours at a moderate hourly cost. The target architecture is three dedicated Hetzner servers, one each for Elasticsearch, MongoDB, and the web stack, plus temporary cloud instances when needed.

The move I expected to dread, Cosmos DB to MongoDB, took two to three days. I had Claude build a query translation layer, then ran Cosmos and the new Mongo layer in parallel with automated validation comparing results until they reached parity. Only then did I cut over.

The other half of the cost story is OCR. Azure Document Intelligence became Tesseract plus my own classifiers that validate output quality, running on an RTX 3090 with 24GB of VRAM sitting in my office, chewing through about 1,500 pages per hour. The office has solar panels that cover both the machine and its cooling, so the marginal energy cost is roughly zero, while the same GPU capacity in the cloud would be prohibitive. Last month the entire Azure bill was down to $1,500, and it keeps falling.

Thirty characters and a running start

The eligibility mismatch that killed my ad conversion also shaped onboarding. At signup I ask for a minimum of 30 characters describing the tenders you care about. The AI turns that free text into monitoring criteria, so your first login already shows relevant tenders instead of an empty dashboard. From there the system watches behavior, thumbs up and down, what you save, where you navigate, and adjusts the filtering, with changes reviewed so nobody gets surprised.

Some users push back on even 30 characters, they want to “just see the app.” But without that input the recommendation quality craters on day one, and day one is when they decide whether to stay. I run two open training sessions every week, offer a short personalized setup meeting for free, and send scheduled emails on days 1, 2, 5, and 10 with practical suggestions instead of forcing a first-time tutorial nobody reads.

The numbers

Today DataDriven is at over 300 paying firms and about 27,000 lei in MRR, roughly 5,000 EUR, up from about 1,500 EUR in September 2025. Total revenue last year was around 40,000 EUR. Growth is a steady, unglamorous 2 to 5 percent per month. The company has been profitable on paper for about three months, with minimal salaries for me and my wife, and Romania had a lesson waiting for me there: you cannot repay money you personally lent the company, or pay dividends, until the balance sheet turns positive. It was my own money, and I still could not touch it.

Pricing is public, no “contact us for a quote,” which in this market is itself a differentiator. The top plan is 290 lei per month and recently gained API access and an MCP server that customers can point Claude or ChatGPT at. The premium tier picked up its first 3 or 4 subscribers within a month of launching, which at this scale is a signal worth respecting.

What is next is more of the same compounding: an MCP server that turns tender documentation into clean, navigable markdown for the AI assistants my customers already use, an end-to-end agent that prepares complete tender submissions, and a slow expansion across more than 60 European procurement portals, with titles translated to Romanian for Romanian firms first. All of it still a company of one, which is how I designed it, and how I intend to keep it for as long as it keeps working.