# AI suggested · accepted
def charge(order):
    key = order.id
    for attempt in range(3):
        res = gateway.charge(order.total, key)
        if res.ok: return res
        wait(2 ** attempt)
    raise PaymentError

# checkpoint · why keys, not retry counts?
# a lost response must never double-charge.

# AI suggested · accepted
def charge(order):
    key = order.id
    for attempt in range(3):
        res = gateway.charge(order.total, key)
        if res.ok: return res
        wait(2 ** attempt)
    raise PaymentError

# checkpoint · why keys, not retry counts?
# a lost response must never double-charge.

# thinking, not vibing
async def rank(applicants):
    scored = await judge(applicants)
    return sorted(scored, key=reasoning)

def cache(rows):
    reads = LRU(rows, ttl=60)
    return reads

# accept the diff, then defend the call

# thinking, not vibing
async def rank(applicants):
    scored = await judge(applicants)
    return sorted(scored, key=reasoning)

def cache(rows):
    reads = LRU(rows, ttl=60)
    return reads

# accept the diff, then defend the call

The hiring evaluation for the AI era

Let them build with AI

Anyone can one-shot working code now, so Candela evaluates the engineering judgement behind every decision a candidate makes.

See In Action Try the live demo

Built by engineers from

College London

What you get

See how each candidate builds with AI

Your applicants build a real project with AI and answer for every decision they accept, so you see the judgement behind the work, not just the code.

Candela Credential 16 applicants

Your applicants

Payment Service · Builders' Project

Maya Rao

Shipped the charge flow and defended every trade-off she accepted.

84/100 Interview

Daniel Okafor

Clean code, but leaned on AI for the hard call without questioning it.

81 Interview

Priya Nair

Shipped it, but let the scope drift past the brief.

Sample assessment · app.candela-ed.com

Why Candela Exists

Old competence signals are killed by AI so, the only way to evaluate engineers is by watching what & how they Build

AI-written

Curriculum vitae

A clean CV is one prompt away.

AI-passed

The take-home

def process(data): results = [] for row in data: if validate(row): results.append(transform(row)) return results def validate(row): return all(f in row for f in FIELDS)

The take-home passes itself.

AI-rehearsed

The interview

Tell me about a hard bug you fixed.

Certainly. In a previous role I diagnosed a race condition by...

How would you scale this service?

Great question. I would begin by profiling...

The answers come rehearsed.

11,000 job applications a minute on LinkedIn, up 45 per cent in a year.

LinkedIn, reported by The New York Times, 2025

Replacing an employee can cost 50 to 200 per cent of their salary.

SHRM, 2022

How this works

Paste the job description, get a Builders' Project

Send the job description, any past assessment, and any rubric you already use. Candela turns them into one live project, and all your applicants build it in our IDE.

Job descriptionpasted from LinkedIn or your ATS

Past assessmentthe take-home you used to send

Your rubrichowever you rate today

Builders' Project Live for all applicants

Payment Service

01Ship the charge flow end to end

02Survive a lost gateway response

03Make a stranger trust the code

Built in the Candela IDE · five dimensions of judgement

Sample project. Yours is drafted from your job description and your bar, and you edit it before it goes live.

The Credential

A Candela Credential for each applicant

Instead of thousands of lines of AI-written code, you get a Candela Credential for each applicant, and the questions to ask the strongest face to face.

Candela Credential Sample

Maya Rao

Payment Service · Builders' Project

84/100

Recommended to interview

How she builds with AI, judged on five dimensions, with the evidence behind every number.

Programming Judgement87

Chose idempotency keys over retry counting, and explained the trade-off unprompted.

Learning Velocity79

First webhook she had written. By the end of the build she was reasoning about delivery order.

Clear Communication84

Your team member can follow the payment flow from the log alone.

Technical Fundamentals82

Caught the AI indexing the wrong column and said why the query plan changed.

Outcome Ownership81

Cut the reporting dashboard to ship the charge flow working end to end.

The interview pack

Three questions, each drawn from a decision in her build.

Ask Maya

You chose idempotency keys over retry counting. Where does that choice break?

From the evidence under Programming Judgement, 87

Ask Maya

How would your webhook handling hold up if event volume grew tenfold?

From the evidence under Learning Velocity, 79

Ask Maya

You cut the reporting dashboard to ship the charge flow. What told you that was the right trade?

From the evidence under Outcome Ownership, 81

The Credential tells you who is worth an interview, and the questions come from the calls they made, so you interview on the substance behind their decisions.

The proof

AI writes the code and the engineer architects

Each accepted AI suggestion opens a question the candidate answers in their own words, so what you see is the engineer behind the AI.

app.candela-ed.com / build / payments.py

payments.py1AI writes the code

12def charge(order):

13 key = order.id # idempotency key

14 for attempt in range(3):

15 res = gateway.charge(order.total, key)

16 if res.ok:

17 return res

2The candidate accepts Acceptedopens checkpoint 11

3They answer in their own wordscheckpoint 11 of 14

You accepted a retry loop around the payment call. What happens if the first attempt succeeded but the response was lost?

The charge carries an idempotency key, so a retry reads the first result back instead of charging twice. I checked the gateway honours the key before I accepted.

Reasoning logged

4The Credential updates

scored from their answers, not their code

Programming Judgement+384

Learning Velocity74

Clear Communication82

Technical Fundamentals79

Outcome Ownership76

Checkpoint 11 of 14 · the full Credential follows the last answer

Sample checkpoint · app.candela-ed.com

Same checkpoints, two candidatesSample

Candidate A41

Candidate B87

Why keep this cache?

“It works and it is faster.”

“Reads outnumber writes 20 to 1 here, and a minute of staleness is fine for a dashboard.”

What breaks first under load?

“Nothing should break.”

“The queue backs up past 5,000 jobs, so I capped it and shed the oldest first.”

Thin answers, weak Credential

Specific answers, strong Credential

Copy-pasting through the project leaves fingerprints in the log.

For engineers

Prove how well you build with AI

Build a Builders' Project, answer for the decisions you make, and let the best teams come to you. Your reasoning is the work sample.

Try the live demo →

Where engineers prove how well they build with AI, and where the best teams evaluate that talent

Bring a role you are hiring for. We will show you the Builders' Project for it, and the Candela Credential you get back.

A 30-minute demo, no prep needed.

Prefer to look first? Try the live demo