The product
What Basanite is, and what it deliberately isn’t.
What is Basanite, in one sentence?
Basanite tests the technical part of the interview, rebuilt for the AI era. Two rounds measure whether a candidate can do the work, in conversation and at a keyboard next to an AI agent.
Why do you say the technical interview is broken?
Two things broke over the last two years. Coding tests turned into a cheating arms race: AI agents and “interview-coder” overlays let candidates pass take-homes and live screens without the skill the test is meant to check. And the skill that now matters, working effectively with AI, isn’t measured anywhere. Ban AI from the interview and you test unaided coding while ignoring how the job is actually done.
What is Basanite not?
We test the technical layer only. Personality tests, motivation profiles, and culture-fit scoring are out of scope. They’re different problems with different evidence and different rules, and bundling them in would weaken all of them. We may build them later as separate products, once the technical layer is solid.
How is this different from coding tests like HackerRank or Codility?
Those tests measure how well you solve isolated puzzles under artificial rules. Basanite measures how you ship complete, well-judged work in a real codebase with an AI agent, the way the job is actually done. Round 2 flips the usual anti-cheating stance: instead of blocking AI, we require it and record how you use it.
How is this different from AI interview platforms like Maki or HireVue?
Most AI interview tools play a fixed set of questions and score the transcript. They don’t adapt to what you say, and they can’t tell real ability from good interview prep. Basanite asks each candidate different questions that map to the same underlying skills and rubrics. Then it adds a second round in a real coding environment, which no transcript-only tool can match.
The two rounds
How the interview itself works.
How long does the assessment take?
Round 1 runs 20–30 minutes. Round 2 is timed by seniority: 35 minutes for junior, 60 for mid, 90 for senior, with an optional 120-minute extension for architecture-heavy senior roles. Both rounds end when the signal is clear, not at a fixed question or task count.
What is Round 1?
A structured conversation. Basanite asks questions grounded in your CV and follows up on anything vague, missing, or unsupported. It picks up the dimensions that come out through how you tell your story: judgment, reasoning, and tacit knowledge.
What is Round 2?
The AI Collaboration Workbench. You get a sandboxed VS Code, a role-matched codebase of several thousand lines, a real ticket set to your seniority, and your choice of AI coding agent. We record keystrokes, agent prompts, git state, and how you verify your work. After the timed session, a 10-minute reflection conversation checks what you did against what you thought you were doing.
What does the codebase look like? Is it a toy?
No. It’s a synthetic project of several thousand lines, matched to the role and seniority: a SaaS codebase for backend SaaS roles, an agentic-systems codebase with retrieval and evaluation harnesses for applied-AI roles, a security codebase with seeded vulnerabilities for security roles. Tickets read like the ones you’d get in your first week. At senior levels they’re deliberately under-specified, so you have to scope the work and sometimes push back on a simulated requester.
If AI use is required in Round 2, how do you prevent cheating?
We flip the usual approach. “Did they use AI” is no longer a cheating vector: we require it and record it. What’s left is someone else doing the work: a third party on the candidate’s machine, or a stand-in for the whole session. We handle that with identity checks at the start, behavioural biometrics sampled through the session and compared to a Round 1 baseline, and a random mid-session check-in where you explain a decision you just made. Real candidates explain it easily from memory. Stand-ins don’t.
What does Round 2 deliberately not measure?
It’s not a hidden algorithm test. There are no LeetCode-style problems. Tickets are ordinary engineering tasks, the kind you’d do any day in the role. The question isn’t whether you can crack a hard puzzle under artificial limits; it’s whether you ship complete, well-judged work the normal way. We also don’t reward using AI “more” or “less”. We look for judicious use: knowing where the agent helps and where it doesn’t.
The eight dimensions
What we score, and why these.
What are the eight dimensions and where do they come from?
Each one has a formal definition, an intellectual source, and a reference list. They draw on cognitive science, philosophy of knowledge, decision theory, organisational psychology, and the new research on human–AI collaboration.
- Judgment Under Ambiguity: committing to a defensible course of action when information is incomplete (Knight; Tetlock).
- Tacit-Knowledge Articulation: surfacing knowledge that lives in practice rather than in text (Polanyi; Nonaka & Takeuchi; Collins).
- Intuition Under Data Scarcity: recognition-primed judgment that distinguishes real expertise from vocabulary (Klein; Dreyfus & Dreyfus; Kahneman & Klein).
- Psychological Safety & Collective Learning: the conditions under which errors surface early and dissent is voiced (Edmondson; Project Aristotle).
- Creative Problem Reframing: recognising when the team is solving the wrong problem (Schön; Dorst).
- Ethical Reasoning in Practice: feeling the weight of real tradeoffs and navigating them with integrity (Aristotle’s phronesis; Rest; AI-ethics applied work).
- Transformative Learning From Experience: updating prior beliefs in proportion to disconfirming evidence (Flavell; Kolb; Mezirow; Argyris & Schön).
- Human–AI Collaboration Intelligence: the calibrated orchestration of AI tooling (Mollick; Dell’Acqua et al.’s “jagged technological frontier”).
Why these dimensions, and not raw coding throughput?
These are the traits that separate strong engineers in complex, AI-era work, and the ones standard technical interviews can’t see. You can’t look them up. They come from real experience and only show up to evaluators who know what to watch for. As AI takes over more of the execution, the human contribution moves toward judgment, synthesis, and collaborative intelligence. Raw throughput is the part AI is replacing fastest.
How do scores get assigned?
Every dimension is scored on a behaviourally-anchored scale. No dimension scores above 3 without a specific quote from the candidate cited as evidence, from the Round 1 transcript, the Round 2 reflection, or an observable pattern in the Round 2 trace. Scores rest on what was actually said and done, not on a general impression.
For hirers
Calibration, reports, and integration.
How does Basanite calibrate to seniority?
Three bands: junior (≈ L3 / 0–3 yrs), mid (≈ L4–L5 / 3–7 yrs), senior (≈ L6+ / 7+ yrs). We don’t try to split sub-bands, say L4 from L5, or Staff from Senior Staff. That’s a job for the final human round, and we don’t claim AI does it well.
Does Basanite recommend hire / no-hire?
No. The decision is yours; we give you better evidence. The hirer report is a briefing for your final human interview: scores backed by quotes, a map of where the candidate has real depth vs surface fluency vs blind spots, and a cheating-risk kept separate from ability.
Which roles and verticals does Basanite support?
25+ representative roles across ten verticals: Consumer Internet & SaaS, Cloud Infrastructure & DevOps, AI / ML / Data, Cybersecurity, Fintech & Financial Services Tech, HealthTech & BioTech, Hardware & Semiconductors, Robotics & Autonomous Systems, Gaming & Interactive Media, and Developer Tools & Languages. Each role and band is its own calibration. The map keeps growing: we add new verticals (quantum software, BCI engineering) and roles (agent-platform engineer) as their markets get big enough to justify dedicated calibration.
Roles outside that map?
Sales, marketing, operations, legal, and executive hiring are out of scope for now. Adding non-technical roles would weaken the technical evaluation. Universities, graduate employers, and training programmes are in scope: their assessment centres cost £500–2,000 per candidate in assessor time, venue, and coordination, and that’s the cost Basanite replaces most cleanly.
Does Basanite integrate with our ATS?
Yes. We connect to Greenhouse, Lever, Ashby, Workday, Bullhorn, and 50+ other ATS providers through Merge.dev. Candidates flow into Basanite automatically when they reach a mapped role, and results push back to their ATS record as a structured note plus a link to the full report PDF. Recruiters stay in their ATS.
How rigorous is the underlying methodology?
Round 1 uses a documented set of 22 named techniques across structure, questioning, depth, consistency, anti-cheating, and scoring, including Narrative Anchoring, Boundary Condition Probing, Counterfactual Pressure, Progressive Excavation, Vagueness Targeting, Honest Failure Elicitation, Predict-Your-Own-Error, Narrative Consistency Tracking, Cognitive Priority Testing, Information Gap Injection, AI Output Signal Detection, Cognitive Load Escalation, Latency Awareness, and Tacit Knowledge Consistency Testing. Round 2 adds six sub-dimensions scored from the trace: Delegation Calibration, Prompt Quality and Decomposition, Verification Rigor, Override Judgment, Engineering Taste, and Solution Completeness. Validation (construct, content, concurrent, and predictive) is ongoing and pre-registered.
For candidates
What the experience is like, and what we do with your data.
Will I be told what is being evaluated?
The methodology is public: the eight dimensions, the design thinking, this page. What we don’t reveal during the interview is which question maps to which dimension. If every question were labelled, you could optimise for the score instead of showing the real signal. What we measure is transparent. Which question measures what is not.
Can I use AI?
In Round 1, no: it’s a conversation, not a coding task. In Round 2, yes, and it’s required. Bring your own tool (Claude Code, Cursor, Copilot, Aider, a local agent). We’re testing whether you ship well-judged work with AI in the loop, not whether you can work without it.
What if I don’t know the answer to a question?
Saying “I don’t know” with real awareness is treated differently from a confident but empty answer. We don’t penalise honest uncertainty. Research on expert judgment treats knowing the limits of your knowledge as a sign of expertise, not a lack of it.
Will I get feedback?
Yes. Every candidate gets a personal feedback report, whatever the outcome. It’s a short, plain-language summary: what you did well, where to develop, and a few concrete suggestions. It’s useful without being reverse-engineerable, so you can’t use it to game a future Basanite assessment.
What about my privacy?
Before any recording starts, a consent screen tells you what we capture (voice and transcript in Round 1; keystrokes, agent dialogue, git state, and time-on-task in Round 2), where it goes (Anthropic, ElevenLabs, Supabase, all listed on our sub-processors page), and how long we keep it (recordings 6 months; transcripts and reports 12 months, then deleted automatically). You can access, export, or erase your data any time at basanite.co.uk/data-rights. We don’t sell candidate data, and we don’t use it to train AI models. Full detail in our Privacy Notice.
Is the interview AI? Can I ask for a human to review my result?
Yes. AI conducts and scores the interview. Under UK GDPR Article 22, you have the right not to be subject to a decision made solely by automated processing. A tickbox on the consent screen, and a self-serve form at basanite.co.uk/data-rights, flag your assessment so the hirer must apply human review before acting on the score.
Still curious
Still have questions?
Register your interest and the team will reach out, or send us a note through the contact page.