Hit.AI Case Study | Muhammed Emir Aydın

Problem

Career tools often give generic advice and opaque scores, leaving candidates guessing why they weren't matched.

Goal

Build an evidence-based AI career assistant that can analyze structured career information reliably and provide useful UI states.

Architecture

Server-side provider selection routes requests to the appropriate AI model, which uses typed tools to return structured data rendered as Generative UI.

Candidate / Job Posting

Next.js Server

AI Provider Selection

Groq / Anthropic / Demo

AI SDK Tools

Structured Result

Engineering Decisions

DECISION

Structured output

WHY

Predictable UI contracts

DECISION

Provider fallback

WHY

App remains usable without one specific provider

DECISION

Server-side provider selection

WHY

Keeps API credentials out of the browser

Key Capabilities

Generative UI

Structured tool outputs are rendered as typed UI rather than untrusted loose text, ensuring reliable visual state.

Job Application Prioritizer Agent

Analyzes candidate profile data against job postings. Inspects requirements, compares via AI, and recommends Apply / Maybe / Skip with confirmed matches, gaps, and next actions.

Reliability Engineering

Handles mid-stream failure states, rate limits, and slow-response states with deterministic sabotage testing, resilience mechanisms, and fallback behavior.

Real Engineering Challenges

Real Engineering Issue

AI SDK v7 Tool Schema Issue

Problem:
Old `parameters` usage caused real Groq tool-calling failure.

Change: Fixed using the newer `tool({ inputSchema, execute })` pattern.

Real Engineering Issue

Next.js Server Action Schema Export

Problem:
Exporting a Zod schema from a `'use server'` file triggered: "A 'use server' file can only export async functions, found object."

Change: Schema definition moved to a pure module outside the server action file.

Guardrails & Verification

Security & Guardrails

Prompt injection protection from job-posting text
Evidence-only candidate claims (no fabricated experience)
No hiring prediction or fake score mechanisms

Testing & Verification

Automated tests (Vitest)
Linting and Type Checking
Manual deployment scenarios

Learnings

AI features still require ordinary software engineering discipline
Structured interfaces make AI features more reliable
Failure states need to be designed intentionally
Testing real integration paths reveals issues mocks miss