Results
Every project below carries one measured number and the stack it ran on. No logos we met once, no anonymous “global enterprise” — these are the four we can talk through in detail, including the parts that were hard.
We name a client only where we have permission to name them.
Web · AI Marketplace
An AI platform for the European car market — connecting buyers with exactly the vehicle they’re after, built for organic growth.
300K+live listings, ingested from many dealer feeds
NLP · Customer messaging
Incoming email and Instagram messages are read, classified and answered from the product and pricing data the team already maintains — a draft per message instead of a blank reply box.
−85%of the manual typing the team did on repeat trading questions
Government · EdTechberon.mon.bg
A public deployment for the Ministry of Education and Science — a digital dictionary published at beron.mon.bg and open for anyone to inspect.
Publicopen at beron.mon.bg — check it yourself
Email Automation · MarTech
End-to-end email-marketing automation — from segmentation to trigger campaigns, integrated across the entire customer-outreach process.
−60%outreach time — the client’s own before-and-after measurement
A project we cannot measure is a project we cannot defend — so we record the baseline before we touch anything.
How every Palamed engagement starts.
Each stage is a fixed fee for a defined scope, agreed before that stage starts. The measured baseline is EUR 6,000–14,000 over two to three weeks, credited in full against a build that starts within three months. A gated pilot on production data is EUR 14,000–38,000 over four to eight weeks; build and handover EUR 30,000–110,000 over eight to twenty weeks; the run stage a rolling retainer of EUR 1,200–4,500 a month. We quote the band before the baseline and a fixed number after it, because the baseline is what makes the number defensible. Every band, and what moves a quote inside it, is on /how-we-work.
The first phase runs two to three weeks. A first automation is usually live for a limited user group four to six weeks after that, assuming your systems expose an API we can write to and someone is authorised to approve exceptions. The slow part is never the model. It is permissions, the absence of a non-production environment, and getting one decision owner named.
You do, on payment: the prompts, the evaluation sets, the pipeline configuration and the infrastructure-as-code. We work inside your repository and your cloud account wherever one exists, and hand over with a runbook. There is no Palamed runtime you have to keep paying for, and nothing load-bearing stays locked inside a tool only we can log into.
Every phase carries a kill criterion written before it starts: the measured result at which we stop rather than extend. If a pilot misses it, we stop, and you keep the baseline measurements, the evaluation set and the written reason. The 2025 MIT NANDA study put the share of GenAI pilots returning nothing measurable at 95%. The gate exists because of that number.
Into systems you control. We run on EU-hosted infrastructure, sign a processor DPA before any access, and use contractual terms that exclude training on your inputs. Sensitive extraction runs on an open-weight model hosted in the EU or on your own hardware. Access is per-person, time-limited and logged, and we keep no copy of your production data after handover.
Because the person who scopes the work also writes it. Palamed is founder-led, with named specialists brought in for the specific project, so there is no pitch team followed by juniors. The honest limit: we run a small number of engagements at once, we do not staff a 24/7 on-call rota, and if you need forty consultants in a room we are the wrong call.
Bring one workflow and the numbers you already have. We will tell you whether automation pays here, and say so plainly if it does not.
A first call takes 30 minutes and produces a written scope, not a proposal deck.