1. Introduction
Applicant Tracking Systems (ATS) are widely used in recruiting to filter, rank, and pre-select candidates before any human review takes place.
These systems rely on keyword matching, structured parsing, and scoring models to evaluate CVs at scale.
This post explores a practical observation:
Hybrid ML profiles (engineering + research + business) can be systematically misranked by ATS systems — even when they are highly relevant.
This post explores a real-world experiment comparing two CV versions and how small structural changes impacted ATS outcomes.
2. Why the rejection surprised me
The position focused heavily on time-series forecasting and logistics-related ML problems - including ARIMA, Prophet, LSTM models, seasonality analysis, and trend forecasting.
I have an interdisciplinary background combining:
- Business Administration
- Software Engineering
- Data Science (certificate program)
- Machine Learning & MLOps in enterprise environments
- Experience across ML systems, MLOps, forecasting, and logistics processes.
In my professional work:
- SAP: ML monitoring and observability systems (time-series, production ML workflows)
- University Hospital Mannheim: drift detection and ML monitoring approaches
- Wayfair: logistics + supply chain integrations, automating data flows (orders, shipments, invoices, inventory)
The role sits exactly at the intersection of engineering, ML, and logistics.
3. How ATS Systems Work (Simplified View)
ATS systems typically perform:
- CV text extraction (PDF parsing / normalisation)
- Keyword matching against job descriptions
- Structured field detection (titles, skills, dates)
- Ranking based on feature signals
Common ranking signals:
- Job title alignment
- Keyword frequency
- Domain consistency
- Career linearity
ATS systems do not evaluate “potential” like a human — they evaluate structured similarity.
4. The Experiment
I created two versions of my CV:
- Version A: Standard CV (context-rich, descriptive)
- Version B: ATS-optimised CV (structured, keyword-focused)
No changes were made to:
- Work experience
- Skills
- Technical background
Only representation changed:
- higher keyword density
- standardised sections
- simplified layout
- explicit skill naming
5. Observation
After submitting the ATS-optimised version, I received an interview invitation.
This suggests:
Small changes in structure and keyword alignment can significantly impact ATS ranking outcomes.
It also raises a broader question:
How much of “candidate quality” is lost in translation through automated filtering?
6. Where ATS Systems Fail
ATS systems struggle with:
- Interdisciplinary profiles
- Non-linear career paths
- Hybrid roles (engineering + research + business)
- Cross-domain experience
Core limitation:
ATS optimises for pattern similarity, not semantic understanding.
This leads to over-reliance on:
- keyword frequency
- title matching
- domain consistency
- linear career assumptions
Result: hybrid profiles may be down-ranked despite strong relevance.
7. AI in Recruiting: Co-Pilot, Not Decision Maker
AI systems are useful for:
- filtering large applicant pools
- extracting structured information
- ranking candidates based on rules
But they should not be final decision systems.
A responsible setup requires:
- human interpretation of rankings
- awareness of system limitations
- context-based evaluation
- ability to override model outputs
AI should assist decision-making — not replace it.
Recruiting contains ambiguity and context that keyword systems cannot fully capture.
8. What Changed in the CV
I did not change my experience - only how it was presented.
Changes:
- more structured layout
- stronger keyword alignment
- clearer skill signalling
- simplified formatting
- explicit technical framing
The key insight:
small changes in wording and structure significantly changed interpretation and made the CV easier to parse for both humans and machines.
9. Key Takeaways
- CV structure directly impacts ATS ranking
- Keyword alignment influences matching more than narrative detail
- Hybrid profiles are structurally disadvantaged
- Machine ranking ≠ candidate quality
- Human review remains essential
10. Discussion
This experiment highlights a tension in recruiting systems:
- scalability requires automation
- automation introduces abstraction loss
Key question:
Have you experienced ATS filtering before?
Do you think your profile was ever misunderstood?
How transparent should recruiting systems be?
Let’s discuss.
Curious how others handle this — especially in ML-driven recruiting systems.
If you're working with ATS systems or hiring pipelines, I’d be interested in how you handle hybrid or non-linear profiles.