Loading the page…
Technical hiring, for the AI era
Hire engineers for how they work with AI.
From the apply link to the offer: an AI screen that follows your plan, a coding round with a real assistant, and an assessment where every score points at the answer that earned it. The AI assesses. A person decides.
EU AI ActGDPRNYC LL144Illinois AIVIAColorado AI Act
Screener
What happens to checkout when that queue backs up?AI fluency
B cites 3 answers
Public apply linksResume read against the roleScreening plans from a job descriptionAI screen, typed or spokenWhiteboard beside the answerCode that the server runsTwo-core assessmentEvery score cites an answerA Cursor-like IDE with an agentPlan · Build · ReviewAuthorship tintingOne scorecard across roundsInvites emailed for youNamed human sign-offAudit log and bias-audit exports
How it works
One candidate, from the apply link to a signed decision.
Five stops, one record. Click through them, or let it play.
NameAmara Okafor
Emailamara@hey.com
amara-okafor-cv.pdf
212 KB · read in 3 s
An AI reads your resume against this role. A person decides what happens next.
- A. Okaforstrong
“Led the move of order intake to an event queue…”
- D. Silvastrong
“Owned Postgres partitioning for 2B rows”
- K. Imaipartial
“Built internal Go services”
- T. Beckerweak
No quote for distributed systems
Invite top 20 to the AI screen
ExperienceSystemsTrade-offsWorking with AI
ScreenerYou mentioned the order queue. What happens to checkout when that queue backs up?
Amara · spokenWe shed load at the edge first, then let intake lag — a late confirmation beats a dropped order…
Fundamentals3.3/4
System reasoning
Trade-off analysis
Communication
AI fluencyB
Solid habits, with gaps worth probing live.
System reasoning · 4
“We shed load at the edge first, then let intake lag…”
Verification habits · 2
“I usually trust the generated tests if they pass.”
Advice: Worth a technical round
A person makes the call.
A person makes the call.
export function rateLimit(max: number, windowMs: number) {
const recent = (hits.get(key) ?? []).filter(t => now - t < windowMs);
if (recent.length >= max) return false;
hits.set(key, [...recent, now]);
return true;
✓ 9 passed · run by the server · 1.4 s
AI screenAdvance
Coding roundStrong yes
Hiring managerYes
HRYes
Average 4.3Lowest verdict Yes
✓ Hired — signed by sarah@yourcompany.com · 24 Sep
1Roles and plans
Paste the job description. Approve the plan.
SynergicHire drafts a screening plan from the role: sections, what each one must cover, a time budget, and where a whiteboard or code belongs. Edit anything. Approving it names you and freezes that version, so every candidate on it was asked the same things.
- Fundamentals and AI fluency both get a section. The plan puts the AI section back if a draft leaves it out.
- Revise it any time. A revision starts the next version. Screens already run keep the one they ran on.
Job description
Senior Backend Engineer, Payments
You'll own order intake end to end — queues, idempotent retries, and the Postgres underneath. We ship with AI agents daily and expect you to check what they write.
Screening plan v1Draft
2The AI screen
A first conversation for every applicant, not just the ten you had time for.
The screener follows your approved plan and probes vague answers. It stops when the plan is covered or its time is spent. Candidates type or speak and can switch at any point. When a section calls for it, a whiteboard or a code workspace opens beside the conversation.
Section 3 of 4 · System designplan v3 · approved
ScreenerLet's design the notification fan-out you described. Sketch the pieces on the board as you go.
Listening…
saved with each answer
The plan stays on the server
The candidate never sees the plan, and nothing they type can change it. Their answers are handled as data, and each reply is checked before it's shown.
Code runs where we can see it
When they send an answer, the server reruns their tests. A pass reported by their browser is thrown away.
Drawings the screener can read
The board is saved with each answer and described to the screener, so it can ask about the diagram itself.
3The assessment
Every score points at the answer that earned it.
Fundamentals and AI fluency are scored separately and never averaged. A score that can't cite a real turn is dropped and reported as a gap. Hover a score to see what it rests on.
Fundamentals3.3 / 4
AI fluencygrade B
System reasoning · 4 Strong
Named the bottleneck, the fix and its failure mode unprompted.
cites 2 answers
Screener
Walk me through the order service you rebuilt. What broke first under load?Candidate
Intake. Synchronous writes to Postgres serialised on one hot row, so p99 went from 80 ms to four seconds at peak.Screener
Sketch what you moved to. What happens when that queue backs up?Candidate
We shed load at the edge first, then let intake lag — a late confirmation beats a dropped order. Idempotency keys make the retries safe. board · 6 shapes
Screener
How do you check code an agent writes for you on work like this?Candidate
I give it the failing test and the one file, read the diff before I accept it, and rerun the suite. If it touches the retry path I write the edge-case test myself.Advice, never a decision
The recommendation reads “Worth a technical round” or “Not on this evidence”. Nothing moves a candidate on its own.
No made-up numbers
If the judge fails twice, the panel says “evaluation unavailable”. An empty panel is better than an invented score about a person.
Only the work
The judge is told to infer nothing about the person beyond what they said and built.
4The coding round
A real project, a real assistant, and a record of how they used it.
A Cursor-like IDE with completions, file-aware chat and an agent. The code is tinted by who wrote it, and the signals update as the session runs. This one is 100 seconds, replayed.
typed by candidateaccepted from AIverified by testsflag for the interviewer
Output
Run or test your code to see output here.
prompt
accept
tests ✗
edit
tests ✓
0:00session · 1:40
AI chat unguarded · Claude
No prompts yet. Reading the tests…
Fluency signals live
Prompt quality—Time to first prompt—AI-generated share0%Verification—Over-reliance flags0Recovery—
Plan
A plan document the assistant is briefed on. A chat reply can become the plan, marked as the AI's.
Build
Monaco, a terminal, tests, and a live preview for React, Next.js, Expo and Node projects.
Review
Their change as a diff against the starting files, with line comments. Offered, never required.
Guarded and unguarded
You decide how much AI they get.
Guarded tests fundamentals with a helper in the room. Unguarded tests how someone ships with a full assistant. The server enforces it, so text pasted into the chat claiming to be a system override unlocks nothing. Try the switch.
unguarded · full assistant
- ChatSyntax, concepts, finding your way round the codebaseavailable
- Full solutions in chatGuarded: refused, redirected to a conceptual hintavailable
- Inline completionsAccept or reject — each one loggedavailable
- Agent modeReads, edits and runs tests behind Allow / Cancelavailable
- Model switchThe configured models, plus Autoavailable
AI-fluency signals
Six things a session shows that a take-home never could.
Each signal comes from timestamped events: edits, prompts, accepted completions, tool calls and test runs. None of it is someone's impression of the transcript. These readouts are from the session replayed above.
0.82
Prompt quality
Specificity, attached context, and how few near-identical re-asks it took.
42s
Time to first prompt
Reasoning attempted before reaching for the AI.
52%
AI-generated share
Share of the final solution's characters that came from accepted AI output.
0.90
Verification
Edits and test runs within 90 seconds of accepting AI output.
0
Over-reliance flags
Blocks over 280 characters accepted and never reviewed or tested.
1.00
Recovery
What happened after the AI was wrong: a fix, a re-prompt, or nothing.
5The decision
Every round on one board. Every outcome signed by a person.
Screening, the coding round, the hiring manager, HR, and any panel or bar raiser you add all live on one record. A round is finished when someone writes it up. The outcome names the person who decided it.
- One scorecard. The lowest verdict sits next to the average, so a mean of 4.1 can't talk you past a strong no.
- Email handled. Invites go out from SynergicHire with replies to you, and interviewers hear when a write-up is owed.
Screening1
A. Weber
Senior frontend
Scheduled · Thu 10:00
Technical2
E. Moreau
Staff engineer
Awaiting feedback
J. Okafor
Senior backend
Invite sent
Hiring manager1
L. Berg
Staff engineer
Scheduled · Fri 15:30
HR1
S. Nakamura
Senior frontend
Scheduled · Mon 11:00
Additional1
T. Novak
Staff engineer
Panel · Wed 09:00
Decision1
R. Diallo
Staff engineer
4 of 4 written up
Closed1
M. Costa
Senior backend
Hired · signed by sarah@
Compliance
Compliant by default, so you can just hire.
The rules for AI in hiring are built into how SynergicHire works. There is nothing extra to set up.
People make the decisions
A named person signs off every round and records every outcome. Nothing is rejected automatically.
Candidates are told first
What an AI reads, what is recorded, and who decides — before they apply, and again before a screen or a round starts.
Audit-ready records
An append-only activity log, selection rates, and a row per decision for an independent bias audit. Exporting is audited too.
Retention and deletion
Recordings are erased on a schedule you set. Candidates can ask to be deleted, and the log keeps who did it and when.
No face, no emotion
No camera-based proctoring and no facial or emotion analysis. Voice audio is kept only if you turn it on.
Run one role end to end. Judge it yourself.
Paste a job description, approve the plan, and share the apply link. You'll have ranked applicants, screen transcripts and cited assessments by the end of the week.