Technical evaluation with evidence
Candidates solve an incomplete case with an AI agent playing a workplace role. Your team receives a report on their questions and decisions, with evidence to prepare the interview.
30 minutes, around a vacancy on your team.
For tech leads and technical recruiting teams.
What is at stake
A good deliverable leaves questions about how it was produced. Answering them candidate by candidate moves the evaluation work back into the interview.
With AI tools available, a correct solution alone says little about the candidate’s questions, assumptions and decisions.
Your team spends another hour discovering how each person reasons. The search depends on finding space in the calendar again.
Adaliza records the conversation and deliverable. You enter the interview with an evaluated case and specific questions to explore.
What your team receives
Each criterion connects a candidate’s decision to the evidence behind it. Go into the interview with specific questions about their work.
Your team defines the rubric. Your team makes the hiring decision.
Open the full real screen (opens in a new tab)Report excerpt · Example data
AI-assisted evaluation
4/ 5Good
Evidence from the conversation
“Can a branch send a file late, or re-send it corrected?”
She asked about late and corrected files before designing the solution. She identified a condition the brief did not spell out.
What Adaliza observes
Your tech lead defines what the agent knows and when it can share it. Candidates talk to that workplace counterpart and build their solution.
Whether they identify constraints, ask for context and check what is missing before committing to a solution.
Which assumptions they state, which alternatives they consider and which risks they accept while information is still incomplete.
How they explain decisions to the role the agent plays and how those decisions appear in the final deliverable.
How it works
Step 1
Your tech lead defines the case and rubric. An assistant proposes a draft; your team reviews and approves it.
Step 2
They receive a link, talk to the agent and submit their solution. Their questions and reasoning are recorded.
Step 3
Review the report by criterion and its supporting excerpts. Prepare the interview and decide who moves forward.
30 minutes, around a vacancy on your team.
Why trust the method
Follow an evaluation from the criterion to its supporting excerpt. That traceability is part of the product we walk you through in a demo.
Your team defines criteria and weights. Inviting freezes the version, so candidates are compared within the same challenge and version.
Scores come with excerpts from the conversation or deliverable. You can review what happened and how the evaluation was justified.
The evaluation supports human review. People’s observations are stored separately from the automatic score.
Developed as an academic project at UTN FRBA.
Meet the teamThey can use AI. We evaluate their questions, decisions and deliverable against the agreed rubric. Evidence comes from the Adaliza conversation and submission; we do not observe their external tools.
They follow an individual link, with no account or installation. They see the time available and submission conditions, and know they are talking to AI and being evaluated.
The scenario, agent role, conditions for revealing information and rubric. An assistant proposes a draft that your team reviews and approves.
A report with criterion-level scores and evidence from the conversation and deliverable. Your team reviews it, adds observations and decides who moves forward.
It is designed for technical roles. Your tech lead adapts the case, agent role and criteria to the vacancy.
It currently works independently in the browser. Your team invites through a link and reads the results in Adaliza; there is no ATS integration.
Name and email identify the candidate; the conversation, deliverable and evaluation form the evidence. Records are kept for at least six months, with controlled access for your organisation’s authorised team.
Illustrative case, built for this page. It is the same one the real platform screens show.
01 · The brief
Farmacias del Sur · Data Engineer, mid-level · 90 minutes
Design and build an Airflow DAG that ingests the daily sales CSVs of 140 branches and the product catalog, and processes them through Bronze, Silver and Gold on PostgreSQL, with data quality checks and a weekly sales and margin report. It must be re-runnable without duplicating data.
What the brief does not say
02 · The conversation
Candidate
Do files always arrive on time? Can a branch send one late, or re-send it corrected?
Mariana · Commercial Analytics, AI agent
Quite often. Some files arrive up to 72 hours late, and corrected ones are re-sent under the same name.
Candidate
How do returns arrive in the file?
Mariana · Commercial Analytics, AI agent
As rows with a negative quantity and the same ticket_id. Summed as they are, they inflate sales.
Candidate
When you say weekly report, does the week run Monday to Sunday?
Mariana · Commercial Analytics, AI agent
No. For us it runs Wednesday to Tuesday: that is when promotion cycles close.
03 · The assumptions
She never asked whether every branch sends the same format, nor whether the catalog changes. The DAG parses one date format and uses the current category, not the one on the sale date.
04 · The deliverable
She submits a link to her repository: the DAG, the models per layer, the checks and a README with her decisions. Before confirming, it is checked to be accessible.
sales-medallion/
05 · The evidence
The evaluation goes through the conversation and the deliverable against the rubric. Soft and technical skills are reported separately.
4.3 / 5
4 / 5 Good
“Can a branch send a file late, or re-send it corrected?”
5 / 5 Outstanding
“Let me confirm: idempotent re-runs, net returns and a Wednesday-to-Tuesday week.”
4 / 5 Good
Partitions by file date to re-run without duplicating, and explains it in the README.
3.9 / 5
3 / 5 Acceptable
Silver checks nulls, types and duplicates, but parses a single date format.
5 / 5 Outstanding
Each task replaces its partition: running the DAG twice gives the same result.
4 / 5 Good
Gold aggregates net sales and margin by branch, category and commercial week.
3 / 5 Acceptable
Picks up late files, but uses the product’s current category.
06 · Comparable
When you invite, the configuration is frozen. You compare candidates on the same rubric, not on impressions.
