Back to Projects
TalentoAI
Full Stack ML ApplicationCompleted

TalentoAI

AI-Powered Interview Preparation Platform

Live Deployment GitHub Repo

The Problem

Candidates lack objective, metrics-driven feedback on interview preparation. Existing tools are heavily rule-based and fail to capture conversational and technical nuance.

Why it matters

Without actionable feedback, candidates fail repeatedly. Objective AI feedback equalizes access to top-tier interview coaching.

Who is affected

Software engineering candidates, fresh graduates, and transitioning professionals.

Architecture & System Design

Frontend

Next.js App Router providing interactive dashboarding for historical performance.

Backend Processing

Serverless functions handling document parsing and orchestrating asynchronous LLM grading tasks.

ML Pipeline

RAG pipeline deployed to retrieve expected technical benchmark answers and evaluate user responses against a technical rigor matrix.

Architectural Reasoning

Chose a highly deterministic prompt-chaining architecture over a single monolithic LLM call to generate segmented metrics across Aptitude, Technical, and Communication.

Alternatives Considered

Considered fine-tuning a BERT-based model for classification but zero-shot Frontier LLMs (like GPT-4) proved superior in nuance detection.

ML & Technical Deep Dive

Model Selection & Training
Core Architecture

GPT-4 API for nuanced semantic evaluation

Training Methodology

Zero-shot Prompt Engineering with highly constrained JSON-schema output requirements.

Dataset

Candidate resume PDFs and textual question/answer transcripts.

Preprocessing Pipeline
  • .01PyPDF2 text extraction
  • .02Regex noise cleaning
  • .03Semantic categorization of domain skills
Evaluation Metrics

92%

ATS Extraction Accuracy

500+

Evaluations Processed
Technical Challenges
Problem: LLM Output Inconsistency
Solution: Forced the LLM onto strict JSON output modes and built a parsing middleware that retries on malformed schema.
Problem: Context Window Limits for Massive Resumes
Solution: Implemented token truncation and hierarchical summarization of legacy work experience.

Core Features

Resume Analytics Engine

Parses resumes to generate an ATS compatibility score and identifies missing critical keywords.

Structured Feedback Matrix

Breaks user answers down by accuracy, clarity, and technical depth.

Progress Tracking

Real-time analytics dashboards to track skill progression over multiple mock interviews.

Results & Impact

Candidates moved from generic 'good job' feedback to hyper-specific 'Your definition of multi-threading lacked mention of race conditions'.

Proven utility in a university environment, helping peers prepare for rigorous tier-1 tech interviews.

15% Increase in User ATS Pass Rates

98% Uptime Architecture

Takeaways & Learnings

What I Learned

Learned the critical importance of constraining LLM outputs (JSON Mode) when bridging AI generation into strict frontend UI components.

Trade-Offs Made

Relied entirely on API calls rather than local inference, tying platform cost directly to usage volume.

Future Improvements

Implement WebRTC for actual live voice-to-voice interview evaluations rather than text-based transcript ingestion.

Tech Stack Foundation

Frontend
  • React
  • Next.js
  • TailwindCSS
Backend
  • REST APIs
  • Node.js
ML / AI
  • OpenAI API
  • PyPDF2
  • Prompt Engineering
Tools
  • Vercel
  • Git

Interested in this architecture?

Let's talk about how I can build something similar for your team.

All ProjectsGet in Touch