how dwij finds evidence.
dwij is an ai agent that hires engineers for startups, and it runs the whole process itself. instead of trusting a cv, it reads the role as concepts, asks each engineer about their real work through situations that unfold as they answer, reads code they wrote when they share it, and reports two things per requirement: do they understand it, and have they done it.
who does the work in dwij?
dwij does: it writes the job post, talks to every candidate, collects cvs, builds and runs each mapping, matches people to the role, builds the evidence cards and sends the shortlist. the founder or hiring manager describes what they are building once and decides who to hire. dwij never rejects anyone. a person at dwij steps in only when something might be unfair to a candidate, such as an integrity flag.
because an agent’s time doesn’t run out at the twentieth cv, every applicant gets a real conversation instead of a six-second skim. and every hire teaches it something: rubrics are recalibrated from which answers predicted a good engineer.
what is an evidence card?
an evidence card is what a founder receives for each candidate. it lists every requirement of the role, marks each one on two questions, understands it and has done it, and shows the source behind every mark: a quote, the stage of a situation they reached, or a line of their code.
each line ends in one of five results: shown, understands, thin, not there, or not assessed. not assessed means the conversation never reached it, which is different from not there. the card comes with an interview guide built from its gaps, and with the candidate’s cv exactly as they uploaded it.
how does dwij read a role without a job description?
the founder describes what they are building, in a chat. dwij turns the product into concepts an engineer must understand, such as idempotency, ledgers or tenant isolation, asks which three or four are deal-breakers, and asks whether a named tool is required or the concept is enough.
the level comes from the work, not a title: will they own a module alone, is their code reviewed, do they make schema decisions. the same list of concepts then drives the job post, the questions, and the card, so nothing is described twice.
what happens in a mapping session?
a mapping is a 20 to 30 minute conversation, in english, by voice or typing. it covers the candidate’s strongest concepts through staged situations, asks whether they have done each thing for real, and ends with a quick round on the rest. it is done once and reused across companies.
a staged situation reveals one stage at a time. a merchant was paid twice; the webhook arrived twice; make the second one harmless. each stage depends on the previous answer, so there is no single question to look up. for a new role, only the concepts not yet covered are asked, in a short role check.
how does dwij use a candidate’s code?
only when the candidate chooses to share a repository. dwij builds its syntax trees and call graphs, traces each function to its commits, and asks only about code the candidate authored. it never runs their code, never shows private code to anyone, and never counts the absence of code against anyone.
code is used to choose what to ask, not to grade the person. a walk-through can be checked against the real call path, and a fix in the history becomes a question about what went wrong. code written with ai help is normal; the conversation shows whether it was understood.
how are answers marked?
against a versioned rubric for each concept that lists what an answer from someone who understands will contain. dwij marks content, never fluency or accent. a model reads each answer; fixed rules decide the mark and the next question. there is no single score.
follow-ups follow fixed rules too: a missed point gets one rephrased question, a hint is recorded, a hypothetical answer gets asked for a real case, and an honest “i don’t know” is accepted. rubrics are recalibrated as hires show which answers mattered.
how does dwij handle cheating?
staged situations, live follow-ups and voice make pasting answers slow and visible. a founder’s interview always includes a question built from the candidate’s mapping, so a stand-in or a pasted answer shows. anything flagged is reviewed by a person at dwij, who talks to the candidate first.
dwij never reads emotion, stress or honesty from a voice or a face, and a machine never accuses anyone: every flag goes to a person first.
what is the ai fluency task?
an optional task in a sandbox, using a coding agent that dwij provides and pays for. the task contains a trap the agent tends to fall into. it shows whether an engineer directs the agent well, checks its work, and catches it when it is wrong.
what does dwij keep to itself?
the rubrics, the thresholds and the models stay private, so they cannot be gamed. what is never private is the source of every line on a card: the founder always sees where a mark came from, and the candidate sees their own card first.