IntelliHire — AI Placement Trainer

Why IntelliHire

🎯

Real Interview Simulation

Each module simulates an actual placement round with adaptive AI — not static Q&A banks. Questions adapt to your ability in real-time using IRT 3PL.

🔵

100% Microsoft Azure

Every service — LLM, speech, database, hosting, CV analysis, CI/CD — runs on Microsoft Azure. Enterprise-grade infrastructure for student placement prep.

🎤

Voice-First Experience

Azure Neural TTS reads questions aloud. Azure STT captures spoken answers. Modules 2, 3, and 4 are fully voice-interactive — just like real campus placements.

👁️

Behavioral AI Analysis

Module 5 uses MediaPipe FaceMesh (468 landmarks), DeepFace emotion detection, and YOLOv8 pose estimation to score eye contact, emotion, and body language.

📊

Scientific Scoring

Module 2 uses IRT 3PL psychometric model — the same algorithm used in GRE/CAT. Maximum Fisher Information ensures every question maximises ability estimation.

🚀

Production-Ready

Not a prototype. FastAPI backend, Beanie ODM, Motor async driver, pydantic v2 schemas, GitHub Actions CI/CD, Azure App Service. Built like a real SaaS product.

5 Modules

Module 01

ATS Resume Analyzer

TF-IDF scoring · Regex NER · Azure OpenAI suggestions · ReportLab PDF

● Live

Module 02

Aptitude Round

IRT 3PL adaptive MCQ · Azure Speech TTS/STT · Fisher Information

◐ Building

Module 03

Technical Round

DSA + LLM voice Q&A · Judge0 sandbox · Monaco editor

◐ Building

Module 04

HR Simulator

STAR NLU · VADER sentiment · RoBERTa NLI · filler detection

◐ Building

Module 05

CV Analysis

MediaPipe FaceMesh · DeepFace emotion · YOLOv8 pose · Confidence Index

◐ Building

AI/ML Algorithms · LLMs · APIs — All 5 Modules

00Master Table

01ATS Resume

02Aptitude

03Technical

04HR Round

05CV Analysis

Module	Algorithm / Model	Type	Library	Purpose
M1	TF-IDF Cosine Similarity	NLP	scikit-learn	Resume vs JD keyword matching
M1	Weighted ATS Scorer (6-component)	Rule ML	Python + textstat	0–100 ATS score
M1	Regex NER Parser	NLP	Python re + NLTK	Extract name, email, skills
M1	Fuzzy String Matching	NLP	rapidfuzz ≥82%	Skill alias detection (JS↔JavaScript)
M1	Azure OpenAI GPT-4o	LLM	Azure OpenAI SDK	Suggestions + bullet enhancement
M2	IRT 3-Parameter Logistic	Psychometric	NumPy + SciPy	Student ability θ estimation
M2	MLE Newton-Raphson	Statistical	SciPy optimize	θ update after each answer
M2	Maximum Fisher Information	Info Theory	NumPy	Optimal next question selection
M2	Naive Bayes Classifier	ML	scikit-learn	Topic domain routing
M2	Azure Speech TTS + STT	Speech AI	Azure Cognitive	Voice MCQ delivery + answer capture
M3	Azure OpenAI Code Analyzer	LLM	Azure OpenAI	Big-O analysis, code review, hints
M3	Domain Router (TF-IDF)	NLP	scikit-learn	DSA/DBMS/OS/OOPs/ML routing
M3	Judge0 Sandbox	Code Exec	Judge0 API	Multi-language code execution
M4	VADER Sentiment Analysis	NLP	NLTK VADER	Emotional tone scoring
M4	STAR Method Detector	NLP	Regex + NLTK	Situation/Task/Action/Result check
M4	Filler Word Detector	NLP	Python re	Count um/uh/like/you know
M4	Azure AI Language (RoBERTa NLI)	Transformer	Azure AI Language	Answer relevance classification
M5	MediaPipe FaceMesh	CV	MediaPipe + TF.js	468-point face landmark detection
M5	DeepFace Emotion Model	Deep Learning	DeepFace + TensorFlow	7-class Ekman emotion detection
M5	YOLOv8 Pose Estimation	CV/DL	Ultralytics YOLO	17-keypoint body language scoring
M5	EAR Eye Aspect Ratio	CV Algorithm	OpenCV + dlib	Eye contact and blink detection
M5	Confidence Index (Ensemble)	ML Ensemble	NumPy + scikit-learn	Weighted 0–100 behavioral score
M5	Azure Video Indexer	Azure AI	Azure Media Services	Cloud emotion + motion analysis

Azure Services Across All Modules

AI / LLM

Azure OpenAI GPT-4oAzure Speech TTS NeuralAzure Speech STTAzure AI Content SafetyAzure AI Language (RoBERTa)Azure Video IndexerAzure Face API

Compute

Azure App Service F1Azure Static Web Apps

Database

Azure Cosmos DB (MongoDB API)Azure Blob Storage

ML / NLP

scikit-learnNLTKNumPySciPytextstatrapidfuzzVADER

CV

OpenCVMediaPipeDeepFaceYOLOv8dlibTensorFlow.js

CI/CD

GitHub Actions (Microsoft)GitHub (Microsoft)

Upload PDF/DOCX

→

Regex NER Parser

→

TF-IDF Matcher

→

6-Component Scorer

→

Azure OpenAI GPT-4o

→

ReportLab PDF

📝

Regex NER + Section Parser

NLP — Entity Extraction

Rule-based Named Entity Recognition using Python regex patterns. Extracts name, email, phone, LinkedIn, GitHub, education, experience, projects. Section detection via 50+ header patterns.

Libraries

Python reNLTKPyMuPDFpython-docx

🔍

TF-IDF Cosine Similarity

NLP — Keyword Matching

TfidfVectorizer on resume + JD. Cosine similarity gives match %. Used for keyword match (30% weight) and experience relevance (20% weight) in ATS score.

scikit-learn TfidfVectorizercosine_similarity

📊

Weighted ATS Scorer

6-Component Composite Score

ATS = Σ(wᵢ × sᵢ) × 100. Components: Keyword Match 30% + Skills 20% + Experience 20% + Formatting 10% + Readability 10% + Sections 10%.

Flesch-Kincaid (textstat)rapidfuzz aliases

🤖

Azure OpenAI GPT-4o

LLM — 3 Prompts

Prompt 1: 8 improvement suggestions. Prompt 2: STAR-method bullet rewriting with metrics. Prompt 3: Professional summary generation. All via Azure OpenAI AsyncAzureOpenAI SDK.

Azure OpenAITemp 0.3 / 0.7

ATS Score Formula

ATS = (0.30×keyword + 0.20×skills + 0.20×exp + 0.10×format + 0.10×readability + 0.10×sections) × 100

API	Provider	Purpose
Azure OpenAI GPT-4o	Microsoft	Suggestions, bullet rewriting, summary generation
Azure Cosmos DB	Microsoft	Store parsed resumes, analyses, JD documents
Azure App Service	Microsoft	FastAPI backend (Python 3.11)
Azure Static Web Apps	Microsoft	Frontend CDN hosting
PyMuPDF + python-docx	Python	PDF and DOCX text extraction
ReportLab	Python	ATS-safe single-column resume PDF generation

θ₀=0

→

Fisher Info → Q

→

Azure TTS

→

Azure STT

→

MLE → θ_new

→

SE(θ)<0.3?

🧠

IRT 3PL Model

P(correct|θ) = c+(1-c)/(1+e^(-a(θ-b)))

θ=ability, b=difficulty, a=discrimination, c=0.25 guessing. S-curve maps ability to correct probability. Core of adaptive engine.

NumPy + SciPyθ: -3 to +3

📐

MLE Newton-Raphson

θ_new = θ + L'(θ)/L''(θ)

Updates θ after every answer by maximizing log-likelihood. Converges in 3–5 iterations. Stops when SE(θ) = 1/√I(θ) < 0.3.

SciPy optimize

🎯

Max Fisher Information

I(θ) = a²[(P-c)²/((1-c)²P(1-P))]

Selects question with highest I(θ) at current ability. Every question maximises information gain — shortest possible test (CAT principle).

NumPy argmax(I)

🎤

Azure Speech TTS + STT

Indian English Neural Voices

TTS: en-IN-NeerjaNeural (female) / en-IN-PrabhatNeural (male). STT: custom phrase list for A/B/C/D recognition. 500K chars + 5hr/month free.

Azure Cognitive ServicesSSML markup

IRT ICC Curve — P(correct|θ) [a=1.5, b=0, c=0.25]

θ=-3 Beginnerθ=0 Averageθ=+3 Expert

Domain

→

GPT-4o Q Gen

→

Azure TTS

→

STT + Code

→

Judge0 Run

→

LLM Eval

💻

Azure OpenAI Code Analyzer

Big-O + Quality + Hints

Reviews submitted code: time complexity, space complexity, correctness vs test cases, edge cases. Returns structured score + improvement hints + optimised version.

Azure OpenAI GPT-4oJSON output

⚙️

Judge0 Code Sandbox

Multi-language Execution

Executes code against hidden test cases. Python, Java, C++, JavaScript. Returns stdout, stderr, execution time, memory. 50 submissions/day free.

Judge0 APIMulti-language

🗺️

Domain Router

TF-IDF Cosine Similarity

Routes to: DSA / DBMS / OS / OOPs / Networks / ML-AI. Cosine similarity between answer and domain knowledge base. Enables domain-specific scoring.

scikit-learn6 domains

🖥️

Monaco Editor

VS Code in Browser (Microsoft)

Microsoft's VS Code editor running in-browser. Syntax highlighting, autocomplete, IntelliSense for Python/Java/C++/JS. Free open-source (MIT).

Monaco EditorMicrosoft (MIT)

HR Persona

→

Azure TTS Q

→

Azure STT

→

STAR + VADER

→

RoBERTa NLI

→

Score Report

⭐

STAR Detector

Situation · Task · Action · Result

Regex + keyword rules detect STAR components. "when I was at…"=S, "I had to…"=T, "I decided…"=A, "which resulted…"=R. Score 0–4 components found.

Python re + NLTK4-component score

😊

VADER Sentiment

Valence Aware Dictionary

NLTK VADER gives compound score (-1 to +1) per sentence. Positive framing → higher score. Detects negativity about past employers (interview red flag). Per-sentence analysis.

NLTK VADERCompound score

🧩

Azure AI Language (RoBERTa NLI)

Natural Language Inference

Azure-managed RoBERTa classifies answer relevance: Entailment (relevant) / Neutral / Contradiction. Scores topic alignment. 5000 records/month free.

Azure AI LanguageNLI 3-class

🚫

Filler Word Detector

um / uh / like / you know

Regex counts: um, uh, like, you know, basically, literally, right?. Filler rate = fillers/words × 100%. >15% reduces communication score. Common interview flag.

Regex counter15% threshold

HR Round Composite Score

HR_Score = (STAR×30 + Sentiment×20 + Relevance×20 + Grammar×15 + Filler×15) / 100

Webcam

→

MediaPipe 468pt

→

EAR Gaze

→

DeepFace Emotion

→

YOLO Pose

→

Confidence Index

👁️

MediaPipe FaceMesh

468 3D Landmarks · 30fps

Detects 468 3D face landmarks in real-time. Tracks gaze direction, head pose (pitch/yaw/roll), eye contact %, smile, micro-expressions. Runs browser-side via TensorFlow.js.

MediaPipe JSTensorFlow.js30fps

👀

EAR — Eye Aspect Ratio

(‖p2-p6‖+‖p3-p5‖)/(2×‖p1-p4‖)

6 eye landmark points. EAR < 0.2 = blink. Blink rate: 12–20/min normal. Too high = anxiety. Too low = staring discomfort. Gaze from iris position.

OpenCV + dlibEAR formula

😮

DeepFace Emotion

7 Ekman Classes — VGG-Face CNN

Detects: angry, disgust, fear, happy, sad, surprise, neutral. Per-frame probabilities averaged over 5-second windows. Micro-expression events (<200ms spikes) flagged.

DeepFaceVGG-Face CNNTensorFlow

🏃

YOLOv8 Pose + Azure Video Indexer

17 Keypoints + Cloud Analysis

YOLOv8-pose: 17 body keypoints for shoulder symmetry, lean, arm crossing. Azure Video Indexer: server-side per-second emotion + motion insights JSON. 10hr/month free.

YOLOv8-poseAzure Video Indexer

Confidence Index Formula

CI = 0.30×eye_contact + 0.25×emotion_stability + 0.25×posture + 0.10×smile_freq + 0.10×blink_norm

Project Roadmap

Development milestones for all 5 modules

Phase 1 · Completed

Module 01 — ATS Resume Analyzer

TF-IDF scoring, Regex NER, Azure OpenAI GPT-4o suggestions, ReportLab ATS-safe PDF. Live and deployed on Azure Static Web Apps + Azure App Service.

Phase 2 · In Progress

Module 02 — Aptitude Round

IRT 3PL adaptive MCQ engine, Azure Speech TTS/STT voice interface, Fisher Information question selection, Naive Bayes topic routing.

Phase 3 · Upcoming

Module 03 — Technical Round

Judge0 code sandbox, domain routing, GPT-4o code review, Monaco editor, difficulty progression logic, voice interview mode.

Phase 4 · Upcoming

Module 04 — HR Simulator

STAR method NLU, VADER sentiment, filler word detection, RoBERTa NLI scoring, GPT-4o HR interviewer persona with adaptive follow-ups.

Phase 5 · Upcoming

Module 05 — CV Analysis + Integration

MediaPipe FaceMesh, DeepFace emotion, YOLOv8 pose, EAR gaze, Confidence Index. Full platform integration and 360° candidate report.

Built By

Student-Cybrarians

B.Tech CSE (AI/ML) — Major Project · 100% Microsoft Azure

⭐ GitHub Repository 🚀 Live Site

Your AI-Powered
Placement Trainer

About IntelliHire

Real Interview Simulation

100% Microsoft Azure

Voice-First Experience

Behavioral AI Analysis

Scientific Scoring

Production-Ready

Your AI-PoweredPlacement Trainer

About IntelliHire

Real Interview Simulation

100% Microsoft Azure

Voice-First Experience

Behavioral AI Analysis

Scientific Scoring

Production-Ready

Your AI-Powered
Placement Trainer