AI-Powered Recruitment Platform
85% faster candidate screening — from 40 hours a week to under 6.
TalentFlow AI — HR Tech
The Challenge
TalentFlow AI was drowning in resumes. Their talent acquisition team spent over 40 hours every week manually screening applications for a growing roster of open positions. The process was slow, inconsistent, and burning out their best recruiters.
Worse, qualified candidates were slipping through the cracks. With each recruiter applying slightly different criteria, match quality varied wildly depending on who reviewed the application. Top talent was accepting offers elsewhere before TalentFlow could even schedule a first interview.
They needed a system that could screen at scale without sacrificing the nuance that good recruiting requires — understanding not just keywords on a resume, but the signal behind them.
Our Approach
We started with a two-week discovery phase, sitting with TalentFlow's recruiters to understand how they actually evaluated candidates — not the formal rubric, but the intuition they'd built over years. We mapped those heuristics into a scoring framework.
From there, we built an ML pipeline that parses resumes using NLP, extracts structured skill and experience data, and scores candidates against weighted job requirements. The model was trained on 18 months of TalentFlow's historical hiring data — including which candidates made it past each stage — so it learned what 'good' looked like for their specific context.
We integrated the system directly into their existing ATS via API, so recruiters didn't need to change their workflow. Candidates are automatically scored and ranked as applications come in.
The Solution
The final platform is a full-stack system: React frontend for the recruiter dashboard, Node.js API layer, and a Python ML pipeline running on AWS Lambda for scalable inference.
The NLP engine handles resume parsing across formats (PDF, DOCX, plain text) with 96% field extraction accuracy. The scoring model uses a gradient-boosted ensemble that weighs skills, experience recency, career trajectory, and role-specific signals.
We built an explainability layer so recruiters can see why a candidate scored high or low — not just a number, but the contributing factors. This was critical for recruiter trust and adoption.
Results
85% reduction in screening time
3x improvement in candidate match quality
Processed 10,000+ resumes in first month
96% field extraction accuracy across resume formats
Recruiter adoption rate of 94% within first two weeks
“Cosmoxyn didn't just build us a tool — they understood how recruiting actually works. The system feels like it was designed by someone who's sat in our chairs.”
Lessons Learned
The biggest lesson was that model accuracy matters less than explainability in hiring. Recruiters won't trust a black box with candidate decisions — they need to see the reasoning. We invested heavily in the explainability layer, and that's what drove the 94% adoption rate. Without it, we'd have built a technically impressive system that nobody used.
Technical Deep Dive
The ML pipeline runs as a three-stage process: (1) Document parsing with Apache Tika + custom NLP for skill extraction, (2) Feature engineering that maps extracted data to job requirement vectors, (3) Gradient-boosted scoring with SHAP values for explainability. The whole pipeline processes a single resume in under 2 seconds. We chose gradient boosting over deep learning because the training dataset (18 months, ~15K applications) was too small for neural approaches to generalize well.
What We'd Do Differently
With what we know now, we'd start with the explainability layer from day one rather than adding it mid-project. We'd also build a feedback loop earlier — letting recruiters flag disagreements with scores — so the model could improve continuously from the start.
Want the technical deep dive?
Read the Behind the Build →Ready to build something intelligent?
Let's discuss how AI can transform your business. Book a free consultation call.