An automated CV filter inherits the bias of the hires it was tuned on, and applies it faster and to more people. We measure that bias against your history before letting it operate, and the system does not make the hiring decision.
3–5 weeks to production · 1 week to measure bias before anything is connected · 3 actionable findings or the diagnostic is not billed
Built on
Ninety seconds: who we are, how we work and what you get at the end.
WhatsApp, the website or email land in the same system: it understands the question, searches your documents, answers with a citation and hands over to a person when unsure.
Example: an employee asks how much holiday is left. The system answers and logs the request.
01 Channels
Searches by meaning and by exact words, and filters by the tag before choosing.
answers with a source · 91 of 100
Swipe to follow the flow →
01
Team hours spent on the most mechanical stage of the process, and good candidates missed through reviewer fatigue.
02
Days pass between applying and a first reply, and the best candidate is already in another process.
03
Coordinating three calendars per candidate is pure calendar work.
04
Leave, employment letters, policies, payroll dates. High volume with answers written down somewhere.
05
Each interviewer asks their own things, and comparing candidates afterwards is impossible.
01
The uncomfortable part
A system that ranks candidates learns from the company’s previous hires. If historically you hired from certain universities, certain neighbourhoods or certain profiles, that is the pattern the system will reproduce — faster, over more people, and wrapped in an appearance of objectivity that makes it harder to argue with. Telling it in the prompt not to discriminate does not change the pattern it learned.
What you can do is measure it. Take the history of applications and hires, run the system as if it were new, and compare the advance rate across groups. A number comes out, and that number gets handed over before anything is connected. If the system would have screened out people the company hired and who did well, it is not ready, and saying so is part of the job.
And the architectural rule: the system ranks and summarises, it does not reject on its own. Every CV stays available, the ranking is a suggestion with its reasoning visible, and the decision is recorded with the name of whoever made it. That is not a legal concession: it is what makes the system defensible when somebody asks why they did not advance.
Against history: whether the share of candidates the system would have advanced is comparable across groups.
What happens if it fails
Proxy signals — university, neighbourhood, photo, name — come out of the context and it gets measured again.
Whether the system would have prioritised the people the company hired and who did well.
What happens if it fails
If it fails here, the criteria do not reflect what the company actually values, and they get rebuilt with the team.
On a sample: whether what the system says about a CV is actually in it.
What happens if it fails
A summary that adds experience the candidate does not have is worse than no summary.
That every answer about policy or benefits comes from the in-force document, cited.
What happens if it fails
A wrong answer about leave or severance ends up as a labour claim.
We do not build systems that reject candidates automatically, or that infer personality traits, emotional state or "culture fit" from a video or a voice. That practice has no evidence behind it and is already restricted in several jurisdictions.
If hiring volume is a handful of people a year, the return is in employee support, not recruitment: less risk, more volume, and answers that are already written down. That is what we recommend starting with in that case.
Applied where there is volume and stable rules, not where there is expectation.
With the evaluation set run before it ever reaches production.
MCP servers against the systems already in operation.
Chunking, reranking and hybrid search, evaluated with recall@k and NDCG.
Key takeaways
The specific risk here is not that the system gets things wrong: it is that it gets them systematically wrong against the same group, faster and at greater scale than a person would.
So the first thing we hand over is not the filter but the bias measurement against your history: if the system would have screened out people the company hired and who did well, it is not ready.
The rule that is not negotiable: the system ranks and summarises, it does not reject on its own. A decision about a person is made by a person, and it gets recorded.
The least glamorous case is the one that pays best: employee support — leave, letters, policies, payroll — high volume, documented answers and zero discrimination risk.
An AI system is worth what its sources are worth. These are the standard connectors; anything with an API or a database connects the same way, and what has no API is handled by file.
SAP
Enterprise ERP
Oracle
ERP and database
NetSuite
Cloud ERP
Salesforce
CRM and service
HubSpot
CRM and marketing
PostgreSQL
Database and pgvector
Microsoft SQL
Database
Snowflake
Data warehouse
BigQuery
Google data warehouse
Databricks
Data platform
Redshift
AWS data warehouse
Synapse
Azure data warehouse
Supabase
Managed Postgres
Workday
Payroll and HR
QuickBooks
Accounting
Sage
Accounting and ERP
Xero
Cloud accounting
Shopify
Catalogue and orders
WooCommerce
Catalogue and orders
Magento
Catalogue and orders
Stripe
Payments and subscriptions
Google Drive
Documents and folders
CSV y Excel
Flat files
Nothing in this category
We built and operated the assistant for a loyalty platform serving more than two million active users across ten countries.
We built the full pipeline: document ingestion and normalization, chunking, embedding generation, vector store on Azure AI Search and Pinecone, and retrieval with grounded generation on LangChain.
We held it above 99% availability for three years.
RAG systems →It gets measured against history before it is allowed to operate. We take past applications and hires, run the system as if they were new, and compare the advance rate across groups, plus check whether it would have prioritised the people the company hired and who did well. Two numbers come out and both are handed over before anything is connected. If the result is bad we say so, because a biased filter applied at scale is a legal and human problem, not a precision detail.
No, and that is an architectural decision. It ranks and summarises: every CV stays available, the ranking comes with its reasoning visible, and a person decides, with their name in the record. This makes the process slower than automatic rejection and makes it defensible, which is what matters the day somebody asks why they did not advance.
We do not build that. Inferring personality, honesty or "culture fit" from a video or a voice has no evidence behind it, and it is already restricted in several jurisdictions. What we do build is interview structure: a consistent guide per role, so two different interviewers assess the same things and candidates can actually be compared afterwards.
Almost always employee support rather than recruitment — even though it is the least glamorous part. Leave, employment letters, policies, payroll dates and concepts are high volume, have written answers and carry zero discrimination risk. It can be running in a few weeks and frees the team from exactly the part nobody wants to do.
Yes, over API or an MCP server. Without integration the assistant explains how to request leave; with integration it says how many days that person has left. We recommend starting with the document layer — which already covers most of the volume — and adding integration afterwards, measuring what it adds before paying for it.
Employee support: three to five weeks to production. Recruitment support takes longer, because it includes the bias measurement against history, which happens first and can change the scope. It is quoted with fixed scope, price and date in a proposal 48 hours after the first call, and the measurement can be contracted on its own as a diagnostic, with its cost credited if you go ahead.
It gets fixed, it gets built, or it is not worth it. And if the diagnostic does not reach three actionable findings, it is not charged.
We use cookies to improve the user experience. Privacy
