Hiring teams were guessing. We built an adaptive assessment engine that scores real skills in real time, and predicts on-the-job success with 94% accuracy.
Fixed-form tests waste a strong candidate's time on easy items and fail to separate weak ones. Our engine runs item-response theory live: after every answer it updates a maximum-likelihood ability estimate and picks the next question with the highest information at that exact level.
The result is a shorter, sharper assessment, convergence in roughly 12 items instead of 40, that produces a calibrated skill vector rather than a raw percentage.
A score is only useful if it predicts the job. We trained a gradient-boosted model on calibrated skill vectors against 18 months of post-hire performance data, retraining nightly on AWS Batch as new outcomes land.
That model outputs a hire-success probability, ranks the candidate pool, and pushes the decision straight into the customer's ATS via webhook, cutting time-to-hire by 40%.
| Component | Technology | Why |
|---|---|---|
| Web app & SSR | Next.js | Streaming UI and server-rendered reports in one runtime. |
| Adaptive engine | Python (IRT) | Mature scientific stack for item-response theory. |
| Question cache | Redis | Sub-ms item lookup keeps each step instant. |
| System of record | PostgreSQL | ACID-immutable attempt and scoring history. |
| Batch & infra | AWS (Batch, ECS) | Elastic nightly retraining and autoscaled serving. |
Each question is chosen in real time from the candidate's current ability estimate.
Raw responses become role-normalized, comparable skill scores.
A gradient-boosted model converts skills into a success probability per role.
Response timing and focus signals flag anomalous attempts for review.
Ranked outcomes advance or reject candidates without manual hand-off.
Every answer and score is stored in PostgreSQL for audit and re-calibration.
“Stop testing what candidates memorized. Start measuring what they can actually do.”
The platform turns hiring from gut-feel screening into a measurable, defensible signal. Adaptive IRT delivers a precise skill read in a dozen questions, a nightly-retrained predictive model maps those skills to a 94% accurate success probability, and ATS webhooks close the loop, cutting time-to-hire by 40% across more than 8,000 assessments a month. Built on Next.js, Python, PostgreSQL, Redis and AWS, it scales with demand while keeping every attempt auditable and every decision grounded in evidence.