Three processes that most companies still run on people and spreadsheets. Headframe builds three AI products for B2B companies that take each of them over end to end: they read, calculate, check and hand back a finished result, predictably and with every step open to inspection. Week to week: fewer hours spent on documents, less money on invoices. Compilore, tothecent.ai, GapRoll. Specs and prices below.
PDF catalogs, manuals, price lists and compatibility tables in one place that answers a salesperson's question in half a minute and shows the page it came from.
For distributors of tools, HVAC and fasteners, and for manufacturers with heavy documentation. Companies where the best salesperson has it all in their head and everyone else calls them. Compilore works from a phone in the warehouse, in Polish and English, and measures which questions it cannot answer. Every week you get a list of what your documentation is missing, before a customer notices.
It doesn't replace your best salesperson. It replaces the phone calls to them.
Product sheet: compilore.pl| Product type | AI knowledge base for companiestechnical documentation search with source citations |
|---|---|
| Time to answer | < 30 sby hand in PDFs about 18 minutes, measured in the pilot |
| Answers with a source | 100%with no source it answers "I don't know" and logs the gap |
| Hallucination on eval | 0.0%test-set score 0.922 |
| Document import | PDF, Word, SharePoint, Google Drive, Dropbox, Confluence |
| Devices | browser + mobile appmobile-first, works in the warehouse and on site |
| Usage analytics | built inanswer success rate, knowledge gaps, who asks what |
| Price | 500 / 1,500 / 3,000 PLN per monthStarter, Team, Business. Pilot on your own documents before any contract |
Assumptions: one manual lookup 18 min, with Compilore 30 s (measured in pilot), 5 days a week, 40-hour week. The sliders are yours; the numbers follow.
It never answers without a source. Every answer carries a citation and page number; if the documentation has no answer, Compilore says so and logs the question as a gap. On technical documentation evals: 0.0% hallucination, 100% of answers cited. A chat that always says something is dangerous in technical distribution, because the customer ends up with the wrong part.
From 500 PLN a month (Starter, one company, 5 users), 1,500 PLN (Team) to 3,000 PLN (Business, SLA). Onboarding starts with a pilot on your documents with a measured result, not with a slide deck.
Documents stay on servers in the European Union (Germany, France) and we train no models on them. Wherever personal data appears in the content, only European models are used.
Knows what every LLM-backed feature really costs, where that money is being wasted, and whether what you get for it is worth it. Matching your OpenAI or Anthropic invoice to the cent comes almost as a side effect.
For teams that have outgrown the question "how much did we spend this month". tothecent plugs into the telemetry you already have and works on three levels. It reconciles: your logs, the provider's usage report and the invoice line up line by line, with caching, batch and retroactive restatements accounted for (providers correct data up to 30 days later). It finds savings: where prompt caching isn't hitting, which workloads qualify for batch, and where a pricier model isn't outperforming a cheaper one, each with a dollar figure attached. It prices value: token cost lands on the feature, team or customer, so the conversation with your CFO starts from margin per feature rather than the invoice total.
Built because we couldn't match our own Anthropic invoice to what the console showed. The first reconciliation ran on our own bills.
Product sheet: tothecent.ai| Product type | FinOps for AI as softwareLLM invoice reconciliation, savings recommendations, cost and value per feature |
|---|---|
| Inputs | 3telemetry via OpenTelemetry, provider usage report, invoice as CSV or PDF |
| Resolution | $0.01hence the name |
| Providers | Anthropic, OpenAI, Mistralprice sheets, caching, batch, retroactive restatements |
| Savings recommendations | caching, batch, model choiceeach with the dollar amount it frees up |
| Cost and value per feature | from tags already in your codeteam, feature, customer: unit economics of every AI feature |
| Prompt content we see | nonetoken counters and cost tags only |
| Alerts | anomalies, spend velocity, budgetbefore the invoice arrives |
| Market and currency | global, USDbilling via Paddle |
| Price | $250 / $900 per monthStarter, Team; Enterprise priced separately |
Because the dashboard, your own logs and the invoice count different things: prompt caching (reads cheaper, writes dearer), batch, retroactive usage restatements (providers correct data up to 30 days later) and price rounding. Those are four different definitions of "usage". tothecent reconciles all three sources line by line and shows the residual with a reason.
No. The wrapper collects token counters and tags only (team, feature, customer). Prompt and completion content never leaves your infrastructure.
The provider console shows a total per API key. tothecent joins that total with your tags (team, feature, customer), with the invoice and with the price sheet, so you see the cost of a single conversation in a given feature, the margin on that feature, and a list of savings with amounts: where caching fails, what can go through batch, where a cheaper model is enough. The console can't compute that because it doesn't know what your feature is.
From the same data we reconcile: cache hit share, the timing distribution of requests (batch candidates), cost versus outcome across models within the same feature. Every recommendation carries the amount and the data it was derived from, so you can check it before acting on it.
A payroll file goes in, an Article 16 pay gap report comes out. In fifteen minutes, with a full audit trail, without storing anyone's salary.
For HR teams at companies with 100+ employees that will have to report their pay gap and answer employees asking about pay ranges once the Polish act is in force. GapRoll has been ready since spring. Until the act arrives we run paygapnews.pl, a Polish-language service tracking exactly what changes in the legislation and when, so HR doesn't find out from a regulation the day before the deadline.
The law is late. We aren't.
Read paygapnews.pl (Polish)| Product type | pay gap reporting toolcompliance with EU Directive 2023/970 for companies with 100+ employees |
|---|---|
| File to report | < 15 minfrom payroll CSV to an Article 16 PDF |
| Input format | CSV from any HR systemexport from Symfonia, Comarch, enova, Excel |
| Audit trail | completeevery calculation can be reproduced for the labour inspectorate |
| Groups under 3 people | suppressedno way to identify an individual salary |
| Data after calculation | deletedin-memory processing, EU servers |
| When in Poland | around Q1 2027the transposition deadline passed on 7 June 2026, progress is tracked on paygapnews.pl |
| Price | 150 / 500 / 1,000 / 3,000 PLN per monthplus a 1,500 PLN one-off for law firms running commissioned audits |
Directive 2023/970 had a transposition deadline of 7 June 2026. The Polish act is in preparation; we expect it around Q1 2027, with the first reporting round for companies with 150–249 employees in mid-2027. Current status, with dates and draft texts, is tracked on paygapnews.pl.
Group employees into categories of work of equal value, calculate the gap in basic pay and in variable components, separately for mean and median, and suppress any group where the result would identify an individual. GapRoll does this from a CSV in under 15 minutes, logging every step.
They are processed in memory on a server in the European Union, the result is written to the report, and the input data is deleted after the calculation. We do not store payroll.
Orders, pilots, questions about catalog items. Tell us what you're after and how many people it affects.