Machine Learning Engineer
1job posting
AI made applying effortless.
It broke hiring.
One click now sends two hundred applications. Nobody can interview that many, so almost nobody gets interviewed at all.
IndusWire interviews every one of them. By voice, in more than 20 languages. Your team opens a ranked shortlist with the evidence behind every score.
1 job posting goes live.
Pune. Twelve openings. Watch what happens to the thousand people who apply.
A thousand in. Twenty-five out.
Every stage named, including the ones you normally never see. Tap or hover any of them - a tap keeps it open. Beside it is the report each of the 390 who finished produces - the real tabs.
| Stage | Candidates | Lost here |
|---|---|---|
| Applications received | 1,000 | - |
| Verification | 900 | 100 |
| CV received | 850 | 50 |
| CV score passed | 550 | 300 |
| Camera, mic and connection ready | 450 | 100 |
| Interview started | 400 | 50 |
| Interview completed | 390 | 10 |
| Shortlisted | 25 | 365 |
Applications received
Someone opened the posting and started applying. The top of the funnel.
reached1,000lost here-
AK
Ananya KulkarniMachine Learning Engineer · Pune · 24 Sep 2026
4.4Interview
C1English CEFR
4.3/5Overall
Scorecard · every row out of 5
Core
Communication skills4.5
Problem-solving4.5
Technical depth4.0
Overall impression4.5
Role-specific
ML fundamentals5.0
Model evaluation & metrics4.5
Data handling & leakage4.0
Production ML & MLOps2.5
Coding assessment4.0
English proficiency
Estimated overall CEFR levelC1
Overall blends the interview at 75% with the coding round at 25% - 4.4 and 4.0 giving 4.3.
RecommendationRecommended
Summary
Strong fundamentals and unusually clear about why a model failed rather than only that it did. Derived the bias–variance trade-off from first principles and caught a target-leakage trap in the follow-up. Thin on serving and monitoring: has trained models but not really owned one in production.
Evidence
“Our AUC looked great until I realised the label was computed
after the event we were predicting.”
11:04 - unprompted leakage catch → Data handling
11:04 - unprompted leakage catch → Data handling
“I’d start with a simple baseline. Half the time it wins,
and then you know the problem is data, not the model.”
06:22 - judgement over novelty → Problem-solving
06:22 - judgement over novelty → Problem-solving
Could not describe how a model
she trained was deployed, versioned or monitored after handover.
19:48 - probed twice → Production ML & MLOps
19:48 - probed twice → Production ML & MLOps
Audio recording
09:41 / 24:18
Video recording
● Rec
▶
09:41 / 24:18
Served over short-lived signed URLs. Clicking any transcript line jumps the player to that second.
Transcript · click a line to play from there
IW
The model is underperforming on the tail. Where do you start?
AK
I’d start with a simple baseline. Half the time it wins, and then you know the problem is data, not the model.
IW
Your AUC was 0.94 in validation and 0.71 live. What happened?
AK
Our AUC looked great until I realised the label was computed after the event we were predicting.
IW
Once that model was trained, how did it get served and monitored?
AK
Honestly, the platform team took it from there. I didn’t own that part.
Problem · Python · 22 of 45 min used
# Stream-safe rolling AUC over a sliding window
def rolling_auc(events, window):
pos, neg, seen = [], [], deque()
for t, score, label in events:
seen.append((t, score, label))
while seen and t - seen[0][0] > window:
seen.popleft()
yield auc([s for _, s, _ in seen],
[l for _, _, l in seen])
AssessmentPass
Weighted into overall at25%
Coding scores · out of 5
Logic & correctness4.5
Code quality4.0
Efficiency3.0
Problem-solving4.5
Edge case handling4.0
Tests
Base cases9 / 9 passed
Edge cases5 / 6 passed
Large inputO(n²) on the window scan
Signals captured during the interview and the coding round
Tab switches
2NotedMoving to another browser tab while a question is open. Counted, with the timestamp of each.
Window focus lost
1NotedThe interview window stopped being the active window - another app, or a second screen.
Large pastes
0ClearText pasted into the code editor in a block, rather than typed. Small pastes are ignored.
Copy events
0ClearContent copied out of the assessment.
Long silences before an answer
0ClearUnusual pauses on questions that do not normally need them.
Device & connection
-ClearLaptop, Chrome 141, stable connection throughout. No mid-interview device change.
How this affects the score: it does not, automatically. None of these signals moves a number on the scorecard. They are surfaced to the person reviewing, with timestamps, so a human decides what they mean. Two tab switches during a system-design question may be a candidate checking their own notes, or their child walking in. We do not pretend to know which, and we do not quietly dock marks for it.
Candidates are told what is recorded before they start, and the same list appears on their consent record.
Screening call · recorded
00:00 / 02:11
Notice period60 days
Expected CTC₹28 LPA
Current locationBengaluru
Willing to relocatePune, yes
Work authorisationIndia, no sponsorship
LanguagesEnglish, Tamil, Hindi
CV match · scored against the role
Role fit4.5
Depth of experience4.0
Production ML exposure2.5
VerdictPassed to interview
Threshold set by employer3.0
ParsedClean · 2 pages
Matched
PyTorch, scikit-learnFound
NLP / transformersFound
Experimentation & A/BFound
Missing
Model serving at scaleNot evidenced
Feature store / pipelinesNot evidenced
Everything recorded against this candidate, newest last
Candidate added · by Priya (recruiter)ok
CV uploadedok
CV score generatedok
Screening link sentok
Screening startedok
Screening completed · 2:11ok
Screening summary generatedok
Interview scheduled · 24 Sep, 10:00 ISTok
Interview reminder sentok
Interview startedok
Coding assessment startedok
Coding assessment completed · 22 minok
Coding assessment evaluatedok
Interview completed · 24:18ok
Interview summary generatedok
Spectator added · hiring managerok
Status changed · Shortlistedok
Note added · “Strong on fundamentals, probe MLOps in round 2”ok
Consent, stored verbatim
I agree that this interview may be recorded, transcribed and assessed automatically, and that my answers and recording may be shared with the employer running this campaign.
accepted 24 Sep 2026, 09:57:12 IST
before the interview began
before the interview began
Shared on LinkedInYes · 25 Sep, 19:04
Data removal requestedNo
Candidate feedback
Was it fair?★★★★★
Would sit anotherYes
“It actually followed up on my answers.
The leakage question was better than most human interviews I’ve had.”
Removal requests go to support@induswire.com, not to the employer.
Everything it does
One platform from the application to the shortlist. Hover to stop the reel.
🎙
Voice interviews
A real conversation, not a form with a microphone. It hears the answer and asks the next question because of it.
📄
CV scoring
Every CV parsed and scored against the role, with the reasoning shown. Unreadable files are flagged, never dropped.
💻
Coding assessment
Generated for the role, run in the browser, scored on correctness, code quality and efficiency separately.
📞
HR screening
Notice period, expected CTC, location, shift flexibility - captured in a recorded call before anyone’s time is booked.
🛡
Integrity signals
Tab switches, focus changes and pasted text, recorded and shown as signals rather than verdicts.
🔗
Campaign links
Publish one link and a whole cohort self-serves. You watch the funnel fill.
🔌
ATS & ERP integrationEnterprise
Push shortlists into your ATS and pull requisitions from your ERP over API and webhooks.
📊
Ranked shortlist
Ordered by evidence, with the timestamp of the moment that earned every score.
📋
Activity log
Every step timestamped - opened, verified, scored, interviewed, reported - on the candidate’s own record.
📜
Consent & feedback
The exact wording each candidate agreed to, stored verbatim, alongside the feedback they left.
🪙
Credits
Pay per module per use. No seats, no minimums, nothing for a campaign nobody applied to.
🌐
20+ languages
Ten of them Indian. The interview and the answers in theirs; your report in yours.
⚙
Role-aware assessments
Paste a job description and the interview, the rubric and the coding task are built from it.
👥
Tenant isolation
Every candidate, campaign and report scoped to one customer at the query layer.
Everything that happens before your team looks
Six things run on every single applicant, whether a thousand apply or four thousand do. None of them cost you an hour.
By the time a name reaches a recruiter it already carries a CV score, a full interview, integrity signals and the evidence behind every number.
- 📄
Every CV is read
Parsed and scored against the role, with the reasoning shown. Duplicates and unreadable files are flagged, never silently dropped.
- 🎙
Everyone is interviewed
By voice, in their language, whenever suits them. It hears the answer and asks the next question because of it.
- 💻
Code is assessed
Generated for the role and scored on correctness, code quality and efficiency - separately, so you can tell them apart.
- 🛡
Integrity is recorded
Tab switches, focus changes and pasted text, shown as signals rather than verdicts. Candidates are told before they start.
- 📊
You get a ranked shortlist
With the timestamp of the moment that earned each score, so you can defend a decision to the person who didn’t get the job.
- ♾
And everyone got heard
Every applicant gets a real conversation and a real answer, whatever you decide afterwards. Nobody else is even trying to do this.
What it changed
Five hiring problems, five different shapes. Pick one.
🏭
Manufacturing
Four plants, one recruiter, and a season that used to take eleven weeks
Operators were hired in seasonal waves. Three recruiters phone-screened for a month before a single technical conversation happened, and the best applicants had usually taken another offer by week three.
Now the posting goes out as one link. Everyone who applies is interviewed the same day they apply, in Marathi or Tamil if that is what they think in, and the plant manager opens a ranked shortlist on Monday morning instead of a spreadsheet of names.
“The shift supervisors stopped arguing with me about who to
call. The evidence was right there under the score.”
Head of Talent, auto components manufacturer
What it moved
11 → 3Weeks to fill a seasonal wave
100%Applicants interviewed, up from 12%
₹9.4LRecruiter hours saved per season
3.1×More candidates reaching a manager
Languages used: Marathi 46%, Hindi 38%, Tamil 11%, English 5%
Regulated roles where the wrong hire is a compliance problem
Every pharmacist has to be assessed on dispensing protocol and patient counselling, not just on a CV. Store managers were doing that themselves, inconsistently, across 210 sites.
The interview now asks the same protocol questions everywhere and scores them against one rubric, with the exact moment that earned each score attached. Audit stopped being a scramble.
“When the regulator asked how we assess counselling skills, we
sent them the rubric and forty recordings.”
Regional HR Lead, pharmacy chain
What it moved
−64%Time to hire per store
−38%90-day attrition
210Stores on one consistent rubric
100%Assessments producible on audit
Languages used: Hindi, Kannada, Telugu, Malayalam, Bengali, English
Four thousand freshers, sixty seats, and three weeks to decide
Campus season is a coordination problem before it is a judgement problem. The old process tested aptitude on a form, then interviewed the top 300 - which meant 3,700 people were ranked on a multiple-choice score nobody would defend in public.
Now every applicant sits a voice interview and a coding assessment generated for the role. Correctness, code quality and efficiency are scored separately, so a candidate who writes working-but-slow code is not confused with one who cannot write it at all.
“We found eleven people in the bottom half of the aptitude
ranking who were obviously worth hiring. That is what we were missing.”
Campus Hiring Manager, IT services firm
What it moved
4,000Interviewed, up from 300
13×More candidates properly assessed
₹1.00Per credit - no per-seat licence
+11Good hires the old ranking missed
Languages used: English 71%, Telugu 12%, Tamil 9%, Hindi 8%
Hiring Arabic-speaking agents when your recruiters do not speak Arabic
The client needed agents fluent enough to de-escalate an angry caller in Gulf Arabic. The screening team in India could assess service instincts but could not judge the language, so every shortlist went through an external assessor and lost four days.
The interview is conducted in Arabic and returns an English report with a CEFR level next to it. The recruiter reads the evidence in their own language and the candidate answers in theirs.
“We stopped paying for language assessors, and the candidates
stopped being judged on an accent in their second language.”
Operations Director, contact centre group
What it moved
−4 daysOff every shortlist cycle
−100%External language-assessor spend
B2+CEFR floor enforced consistently
2.4×Offer-accept rate vs the old process
Languages used: Arabic 68%, English 32% - every report in English
Seasonal retail, where nobody waits a week for an interview slot
Store staff apply to several chains at once and take whoever calls first. A two-day gap between application and first contact was costing the best applicants outright.
Candidates now interview on their phone within minutes of applying, in Spanish, at whatever hour suits them. Nearly half of the interviews happen outside office hours - which is exactly when someone working a current retail shift is free.
“Our best hires came in at eleven at night. We were never going
to reach those people with a scheduled call.”
Head of Retail Operations, LATAM chain
What it moved
<10 minApply to interview, median
47%Interviews outside office hours
−29%Drop-off between apply and assess
+18%Offer-accept rate
Languages used: Spanish 94%, English 6%
Composite examples drawn from real deployment patterns. Named references are available under NDA - ask us.
“Four thousand applications. Sixty openings. Three people on my team.”
That is the job, every season. The coordination is the entire cost of volume hiring - and it’s paid before anyone knows which candidates were worth the time.
One link. A whole cohort.
Publish it once. Candidates verify themselves, upload a CV and sit the interview when it suits them. You watch the funnel fill and open only what’s worth opening.
20+Languages, ten of them Indian
100%Of applicants interviewed
₹1Per credit - no seats, no minimums
You applied to two hundred jobs.
Nobody called.
It was never about you - nobody had time to read them. Sit the same assessment employers actually run, get a real report on where you’re strong and where you’re not, and let the companies hiring right now see it.
Opening soon. About what a coffee costs.
You pay for interviews that happen
No seats. No minimums. No annual lock-in. A campaign that nobody applies to costs you nothing at all.
Start free
500 credits
On the house when you sign up. No card, no call, no trial clock.
Worth20 CV scorings
ora screening call + a coding test
Full reportsIncluded
Pay as you go
₹1 / credit
One credit is one rupee. Each module is priced per use, so you only buy the steps you actually run.
CV scoring25 credits / CV read
HR screening100 credits / call
Voice interview40 credits / minute
Coding assessment200 credits / assessment
Enterprise
Custom
For volume hiring across sites, with procurement and security review.
Data residencyYour region
ATS & ERP integrationIncluded
SSO & role controlsIncluded
Dedicated supportNamed contact
Full pricing, worked examples and FAQ →
One credit is one rupee (one US cent outside India), so a
twenty-minute voice interview is 800 credits and reading a CV is 25. You are charged for what
runs - a campaign nobody applies to costs nothing at all.
Security & data protection
Hiring data is among the most sensitive a company holds - names, CVs, recordings, salary expectations. It is handled accordingly, and a recruiter can always produce the record when someone asks for it.
GDPR EU & UK
DPDP Act 2023 India
EEOC US hiring fairness
NYC Local Law 144 automated hiring tools
ISO 27001 & SOC 2 controls roadmap
🔒 Encrypted throughout
TLS in transit, encryption at rest for CVs, recordings and transcripts. Media is served through short-lived signed URLs, never a public bucket.
🌐 Data residency
Choose the region your candidate data lives in. Cloud only - a single managed platform we patch and monitor, rather than a copy running unattended in your estate.
👤 Tenant isolation
Every customer’s candidates, campaigns and reports are scoped to that customer at the query layer. One employer can never walk a candidate id into another’s records.
📜 Consent, stored verbatim
The exact wording a candidate agreed to is stored with its timestamp - not a reference to a policy that may since have changed. Producible on request.
🗑 Right to erasure
Candidates ask us, not the employer, and removal requests go to support@induswire.com. Retention is bounded and deletion is honoured across recordings, transcripts and reports.
🔍 Purpose limitation
A candidate’s data is used to assess them for the role they applied to. It is not sold, not used to train a general model, and not shared between employers.
⚖ Decisions you can defend
Every score points at the moment in the interview that earned it. When a candidate or a regulator asks why, there is an answer that is not “the model said so”.
📋 Full audit trail
Every step is timestamped - link opened, verified, scored, interviewed, reported, consent given, feedback left - and visible on the candidate’s own record.
Processing agreements, sub-processor lists and our security documentation are available to customers and prospects - ask us.
Questions
Does the candidate talk to a person?
Not in the first round - IndusWire conducts it, and your team meets the candidates who come through. Candidates are told this before they start.
Which languages?
Over twenty, ten of them Indian. The interview, the follow-ups and the answers are all in the candidate’s language; your report is in yours.
How do you stop cheating?
Tab switches, window focus changes and pasted text are recorded and shown as signals, not verdicts - because none of them individually proves anything. You decide what they mean.
What if a candidate’s CV won’t parse?
It’s flagged on the funnel as a named failure rather than dropped. The same is true of every other drop-off: each one tells you which step is leaking.
Where is our data held?
In the region you choose, on our managed cloud. Every customer is isolated at the query layer and deletion is honoured on request. We do not offer on-premise deployment.
How much does an interview cost?
Credits start at ₹1, and CV scoring, screening, interviews and coding assessments are each priced per use. You pay for what runs.
Let everyone
be heard.
Tell us what you’re hiring for and we’ll send you the interview your candidates would actually sit.
Or write to support@induswire.com - Pune, India.