AI / LLM architecture

Sophia calls OpenAI for resume field extraction and interview reports. Live interview dialogue runs in an external bot; the API stores transcript and metrics but does not run an in-session LLM.

flowchart TB
  Resume["Resume upload"] --> Parse["Text extract + LLM parse"]
  Parse --> Candidate["Candidate form / save"]
  Interview["Interview complete + settle"] --> ReportGen["Report generation worker"]
  ReportGen --> Report["Interview report"]
  Candidate --> Snap["Snapshot at schedule"]
  Snap --> Bot["External bot interview"]
  Bot --> Interview
  Snap --> ReportGen
  Bot --> ReportGen

Models

Use case	Model	Purpose
Resume extraction	`gpt-5.4-mini`	Structured candidate fields from resume text
Resume extraction (fallback)	Gemini (`gemini-2.5-flash-lite` default)	Retry on primary rate limit / server errors
Interview report	`gpt-4.1-mini`	Scored evaluation from snapshot + transcript
Insufficient evidence	— (deterministic)	Short interview or too few answers; no LLM call

Both LLM paths use strict JSON schema outputs validated with Zod. Reports use a dedicated OpenAI client (no Gemini fallback).

AI / LLM architecture ​

Models ​

AI / LLM architecture

Models