Case study - product

Parci: the real-estate AI platform Digiton built and runs

Parci is our own product. It answers one question about Portuguese and Estonian property: what can legally be built on this plot. You point it at a location, and about 47 seconds later you get a cited feasibility report. Digiton built the whole thing, launched it in 2026, and runs it in production today.

Planning rules nobody can read

Portugal has 308 municipalities. Each publishes a PDM, the Plano Diretor Municipal, and that document decides what you can build on any given parcel. The rules are real and enforceable. The format is hostile. Zoning maps sit on WMS servers built for GIS software, and the written regulations live in PDFs published through the Diario da Republica, many of them scanned, some running past 200 pages.

So a developer weighing a plot in Braga has two options. Hire a lawyer or an architect and wait weeks for a preliminary opinion. Or guess. Both cost money, and guessing costs more.

Zoning is only the first layer. A parcel can sit inside a REN or RAN reserve, an ARU, or the protection zone of a classified building, and each overlay comes from a different institution with a different map. DGT publishes some layers, ICNF others. Stacking them by hand is an afternoon of GIS work per parcel, which is why almost nobody does it before making an offer.

The frustrating part is that the data existed the whole time. Every building index and every REN or RAN restriction was published somewhere official. Nobody had wired it into a system a normal person could ask a question.

What we built

The product splits into a data pipeline, an agent graph on top of it, and a report engine at the end. Each earned its own share of scar tissue.

The data pipeline

We ingested PDM zoning data for more than 300 municipalities across Portugal and Estonia. Roughly 1,400 zone rows, each tied to real geometry, resolving down to the freguesia. On top of that sits an index of 1,240 legal citations pulled from the Diario da Republica, plus heritage and reserve layers from DGT and ICNF. When Parci cites a rule, it points at the official source.

The zoning layer starts from the municipalities' own PDM shapefiles, matched against the written regulations they implement. Municipality market data refreshes weekly, so the financial side of a report reflects current conditions rather than a snapshot from whenever the pipeline first ran.

Estonia came second. Different registry, different language, same structure once ingested, which is how we knew the pipeline travels.

The six-agent graph

The reasoning layer is a graph of six agents, documented in full in our public deep-dive post. One agent works out what the user is actually asking. One retrieves the relevant legal and planning documents through hybrid RAG, meaning keyword matching combined with semantic search. A third cross-checks the retrieved material against that specific municipality's own regulations, because a rule that holds in Lisbon can fail 20 kilometres away. Then a drafting agent writes the analysis, a verification agent checks every claim against the sources it cites, and an assembly agent formats the result for someone who has never read a planning code.

The verification step is the point. In a domain where a wrong answer creates legal liability, evaluation gates are load-bearing. Every answer gets checked against its sources before it reaches the user, because a confident mistake about what the law allows is worse than no answer at all. Speed came afterwards. The graph originally ran sequentially. We refactored it so independent stages run in parallel, and that is how a full municipal analysis lands in around 47 seconds without cutting a single gate.

The report engine

The output is a 14-page PDF. It covers zoning classification for the exact parcel, compliance checks against RJUE, RGEU, SCIE and SCE, plus accessibility and parking rules. Three massing options come with a pro-forma behind each, including IRR, IMT and stamp duty at 2026 tax brackets, alongside hourly shadow studies across all 12 months. The report also names the permit path: whether the project needs a PIP, a licenca, or a comunicacao previa. There is a one-slide deck for investor conversations and shareable links for teams.

Two more pieces round out the product. Parci renders each parcel in real 3D city context, so the massing options sit among the actual neighbouring buildings instead of floating on a white grid. And an analysis can start from a listing instead of an address: paste a URL from Idealista, Imovirtual, Casa Sapo or Custojusto and the engine runs against that property. Finished reports share as public links, with optional watermarking for teams that care where their analysis ends up.

Every report carries the same caveat in plain sight: this is a preliminary, informative analysis. A licensed architect signs the real project. Selling certainty you cannot deliver is how platforms in this space die.

The numbers

Coverage stands at more than 300 municipalities across Portugal and Estonia. The citation index holds 1,240 legal references. The zoning layer carries roughly 1,400 PDM rows. A full analysis takes about 47 seconds and lands as a 14-page cited PDF.

The comparison that matters is against the status quo. A preliminary opinion that used to take weeks now takes under a minute, and it arrives with its citations attached, each one traceable back to the Diario da Republica. That gap, weeks to 47 seconds, is the whole business case.

Plans run from 29 to 499 euros per month. One-off checks cost 49 euros per report, or 119 euros for a pack of three. A Studio seat is 149 euros per user per month with 10 reports included. The Developer tier, at 499 euros per month, carries unlimited reports and 3 users. Municipalities and other public bodies get custom pricing, because a camara evaluating its own territory is a different conversation from a developer evaluating a deal. A free tier with 3 analyses exists so anyone can test the engine on a real plot before paying a cent.

What Parci proves about our platform work

Plenty of agencies talk about building AI products. Parci is ours, live at parci.eu since 2026, and it exercises every discipline we sell to clients. The data work alone spans WMS endpoints, scanned PDFs, and national registries in two languages. The AI layer is a six-agent graph with real evaluation gates rather than a single prompt in a trench coat. And the platform runs as a business, with billing from 29 to 499 euros a month, usage limits, and the unglamorous operational work that decides whether software survives contact with paying users.

There is no client to hide behind here. If a citation is wrong or a report stalls, that is our problem, in production, in front of paying users. That pressure is worth more than any portfolio page, and it is why we point prospects at Parci before we show them anything else.

Running our own product also changes how we build for clients. The sequential-to-parallel refactor that took analysis time down to 47 seconds taught us more about agent orchestration than any tutorial could. The weekly market data refresh taught us what data maintenance actually costs once a pipeline covers 300 municipalities. Lessons like these come from operating software, and they show up in every platform we scope for someone else.

When a company asks Digiton to build a platform, this is the reference. We built it and we still carry the pager. If your industry has its own version of the PDM problem, official data that answers nothing because nobody can read it, that is exactly the kind of system we build.

Frequently asked questions

Related