SellYourLogic · AI Interview Coach
Make a model return data, not prose
An AI coach writes interview questions from a resume and grades every answer. The model always replies in a fixed JSON shape that code can check, and every score is computed by code.
The business need
An expert practises a realistic interview built from their resume and the job they want, then gets honest, consistent grades they can trust from one session to the next.
0.90
similarity at which two bank questions count as duplicates
5
rubric criteria per answer, each 0 to 100
2 min
a failing provider sits out before it is tried again
How one request flows
From the visitor's words to what they see, step by step.
-
1
Resume and job
The resume is turned into text once; the job description is optional.
-
2
Generate
The model writes questions in a fixed JSON shape (a schema).
-
3
Validate
Code checks the shape. A broken reply is retried, never stored.
-
4
Grade
One answer at a time, against a 5-criteria rubric, in the background.
-
5
Compute
PHP adds up every total and the readiness score.
-
6
Report
The model writes only the words around the numbers.
The key ideas
Schemas, not free text
Every model call asks for a fixed JSON shape. Code validates it, so a reply always becomes a typed object or fails in a way we can retry.
Numbers never come from the model
The model gives one grade per criterion. Totals, averages and the readiness score are computed by code, so a report is the same whichever model graded it.
A grade must be complete
A reply missing any rubric score is rejected and retried with the next provider. It is never saved as a zero, which would silently punish the candidate.
Provider chain with a bench
An ordered list of AI providers. One that fails is "benched" for two minutes and the same request moves to the next, like a simple circuit breaker.
Slow work in a queue
Generating and grading take seconds to minutes, so they run as background jobs. The page polls, and timers are kept by the server.
Calibration with a gold set
Answers graded by people are shown to the model as examples (one strong, one weak) inside every grading prompt.
Embeddings for a question bank
Good questions are stored with a vector of their meaning, so near-duplicates are removed and questions can be ranked against a job description.
Graceful offline mode
With no AI, bank interviews still run with instant rule-based checks and a provisional report that upgrades itself when the AI returns.
What testing taught us
Found by running the real model, not by guessing. This is the most useful part.
"This model feels better" is not evidence
We measure a model against the human-graded set before switching: average gap to the human score, bias (is it too generous?), agreement within 10 points, and correlation.
A missing score is not a zero
Early on, an incomplete reply would have been stored as 0 for that criterion. Now it is a retryable failure, so a candidate is never graded down by a model error.
Duplicates hide behind new words
A paraphrased question scored 0.92 similarity with the original, while the closest truly different pair scored 0.865. So 0.90 became the line for "same question".
What you learn at this level
- Prompts that return data
- JSON schemas and validation
- Retries and fallback
- Embeddings for de-duplication
- Measuring a model with a gold set
Where it stops
This level has no retrieval and no agent: every call is one prompt and one structured answer. That is exactly what the next levels add.
AI Playground
Change chunk size, similarity threshold, model and prompt on our case studies, and see what was retrieved, the exact prompt, the answer, the speed and the cost.
Open the Playground →Live workshops
Hands-on sessions that build each level step by step, hosted by our founder and by experts on SellYourLogic.