🎯 ARIA — AI Readiness Interview Agent

Your Personal AI-Powered Mock Interview Coach

IBM SkillsBuild × AICTE Internship 2026 — Problem Statement #22

---

"Don't just prepare for interviews — simulate them with AI."

---

📌 Table of Contents

• About ARIA
• Demo
• Features
• Architecture
• Tech Stack
• AI Tools Used in Building ARIA
• Installation
• Deployment
• Project Structure
• Token Optimisation
• Resume Feature

---

🤖 About ARIA

ARIA (AI Readiness Interview Agent) is an intelligent conversational agent that conducts full end-to-end mock interviews for job seekers targeting top IT companies like TCS, Infosys, Wipro, Accenture, and Cognizant.

Unlike generic Q&A tools, ARIA runs a complete structured interview across three rounds, scores every answer in real-time, adapts difficulty as you progress, and hands you a personalised Final Report Card with a 7-day study roadmap — all powered by Meta Llama 3.3 70B Instruct via IBM watsonx.ai.

Now with Resume-Aware Interviews — upload your PDF/DOCX resume once and ARIA will ask questions directly tied to your real projects, skills, and achievements.

---

🎬 Demo

🎥 Demo Video

---

✨ Features

Feature	Description
🎯 Role-Based Questions	Tailored questions for any IT role and target company
📄 Resume-Aware Interviews	Upload your PDF/DOCX/TXT resume — ARIA asks about your real projects, skills & achievements
📊 Real-Time Evaluation	Score (X/10) + concise feedback after every answer
💡 Confidence Tracking	Measures confidence growth from start to finish
🔄 Adaptive Difficulty	Questions get progressively harder as you perform better
🎙️ Voice Input	Speak your answers using your microphone (Chrome/Edge)
📋 Final Report Card	Overall scores, top strengths, improvement areas + Resume Alignment score
🗺️ 7-Day Study Roadmap	Personalised preparation plan based on your performance
💬 Multi-Turn Chat	Natural conversational interview flow — no forms or dropdowns
🏢 Multi-Company Support	TCS, Infosys, Wipro, Accenture, Cognizant and more
⚡ Token-Optimised	Server-side session store — flat token cost regardless of interview length

---

🏗️ Architecture

┌──────────────────────────────────────────────────────────┐
│                    User (Browser)                        │
│           Chrome / Edge  (voice input supported)         │
└─────────────────────────┬────────────────────────────────┘
                          │
┌─────────────────────────▼────────────────────────────────┐
│           Streamlit Frontend  (app.py · :8501)           │
│                                                          │
│  • Chat UI — displays messages                           │
│  • Voice widget — Web Speech API (browser-native STT)    │
│  • Resume uploader — PDF / DOCX / TXT (optional)         │
│  • Sends only: { session_id, message }                   │
│    ↳ NO conversation history in the HTTP body            │
└─────────────────────────┬────────────────────────────────┘
                          │  POST /chat        { session_id, message }
                          │  POST /session/new
                          │  POST /resume/upload { session_id, file }
┌─────────────────────────▼────────────────────────────────┐
│           FastAPI Backend  (main.py · :8000)             │
│                                                          │
│  • Server-side session store  (_SESSIONS dict)           │
│    ↳ { history: [...], resume_digest: str | None }       │
│  • Resume parser — PyMuPDF (PDF) + python-docx (DOCX)    │
│  • One-shot digest call → compact ~120-word summary      │
│  • History trimming + dynamic token budget               │
│  • ARIA system prompt engine (digest injected when set)  │
│  • Background model warmup at startup                    │
└─────────────────────────┬────────────────────────────────┘
                          │  ibm-watsonx-ai SDK
┌─────────────────────────▼────────────────────────────────┐
│           IBM watsonx.ai                                 │
│           Meta Llama 3.3 70B Instruct                    │
│           (Free Lite tier — Dallas region)               │
└──────────────────────────────────────────────────────────┘

Key design decision — server-side sessions: The frontend never sends conversation history. Only session_id + message travels over HTTP. The backend holds all history, trims it to a rolling window, and sends a fixed-size context to the LLM — keeping input token cost flat at every turn of the interview.

Resume design decision — one-shot digest: The raw resume is never stored or re-sent. One LLM call converts it to a ~120-word structured digest on upload. That digest (~130 tokens) is injected into the system prompt on every subsequent call — keeping the per-call overhead flat and the raw file off the wire entirely.

---

🛠️ Tech Stack

Layer	Technology	Purpose
LLM	Meta Llama 3.3 70B Instruct	Interview agent brain — question generation, scoring, report card
LLM Platform	IBM watsonx.ai	Hosted inference, API access, token usage tracking
Backend	FastAPI + Uvicorn	REST API, session management, prompt engineering
Frontend	Streamlit	Chat UI, sidebar controls, real-time display
Voice Input	Browser Web Speech API	Microphone-to-text (zero external dependencies)
Resume Parsing	PyMuPDF + python-docx	Extract text from PDF and DOCX resumes
SDK	ibm-watsonx-ai	Official Python SDK for watsonx.ai model calls
Language	Python 3.10+	Core language
HTTP Client	requests	Frontend → Backend API calls

---

🤝 AI Tools Used in Building ARIA

ARIA was built using a combination of human engineering and AI-assisted development:

Tool	Role in This Project
IBM watsonx.ai	Hosts Meta Llama 3.3 70B — the LLM that powers ARIA's interview agent at runtime
IBM Bob (AI Software Engineer)	Used throughout development to architect, debug, and optimise the entire codebase — from designing the server-side session store and token budget system, to fixing prompt behaviour bugs, implementing the voice input feature, and writing all deployment configuration
Meta Llama 3.3 70B Instruct	The foundation model ARIA uses to conduct interviews, score answers, and generate report cards

ARIA was designed and developed with IBM Bob as the primary engineering assistant — responsible for the full architecture, all code optimisations, the voice feature, and the token-reduction work that brought input cost from ~1,400 tokens/call down to ~505 tokens/call.

---

⚙️ Installation (Local)

Prerequisites

• Python 3.10+
• IBM Cloud account — free at cloud.ibm.com
• IBM watsonx.ai project + API key
• Chrome or Edge browser (for voice input)

Steps

# 1. Clone the repository
git clone https://github.com/yourusername/interview-trainer.git
cd interview-trainer

# 2. Create virtual environment
python -m venv ml
ml\Scripts\activate        # Windows
source ml/bin/activate     # Mac/Linux

# 3. Install dependencies
pip install -r requirements.txt

# If upgrading an existing install, add the two new resume packages:
pip install pymupdf==1.24.9 python-docx==1.1.2

# 4. Create .env from template
copy .env.example .env     # Windows
cp .env.example .env       # Mac/Linux
# Open .env and fill in your IBM credentials

Environment Variables (.env)

IBM_API_KEY=your_ibm_cloud_api_key
WATSONX_PROJECT_ID=your_watsonx_project_id
WATSONX_URL=https://us-south.ml.cloud.ibm.com

Get your credentials at:

• API Key → cloud.ibm.com/iam/apikeys
• Project ID → watsonx.ai dashboard → your project → Manage tab

---

🚀 Running Locally

Open two terminals (both with ml\Scripts\activate):

Terminal 1 — Backend:

uvicorn main:app --reload

✅ Wait for: Application startup complete.

Terminal 2 — Frontend:

streamlit run app.py

✅ Opens at http://localhost:8501

Open in Chrome or Edge for voice input support.

---

☁️ Deployment

ARIA has two services — deploy them separately:

Service	File	Recommended Platform
FastAPI backend	main.py	Render (free)
Streamlit frontend	app.py	Streamlit Community Cloud (free)

Render (Backend)

1. New Web Service → connect GitHub repo
2. Build: pip install -r requirements.txt
3. Start: uvicorn main:app --host 0.0.0.0 --port $PORT
4. Add env vars: IBM_API_KEY, WATSONX_PROJECT_ID, WATSONX_URL
5. Copy your service URL: https://aria-api-xxxx.onrender.com

Streamlit Community Cloud (Frontend)

1. New app → select repo → main file: app.py
2. Settings → Secrets → add:

API_URL = "https://aria-api-xxxx.onrender.com"

---

📁 Project Structure

interview-trainer/
│
├── main.py                  # FastAPI backend — ARIA agent, session store, resume upload, LLM calls
├── app.py                   # Streamlit frontend — chat UI, voice input, resume uploader
├── watsonx_client.py        # Standalone IBM watsonx.ai smoke-test helper
├── prompt_templates.py      # Prompt builders — resume digest, interview questions, difficulty
├── sample_prompts.py        # Pre-written test conversations for 5 companies/roles
│
├── good_interview.gif       # Demo — high-scoring interview run
├── bad_interview.gif        # Demo — weak answers with ARIA's feedback
│
├── requirements.txt         # Python dependencies (pinned)
├── Procfile                 # Process file for Render / Railway / Heroku
├── .env.example             # Credentials template — safe to commit
├── .env                     # Your live credentials — NEVER commit this
├── .gitignore               # Excludes .env, ml/, videos, logs, uploaded resumes
└── README.md                # This file

---

⚡ Token Optimisation

ARIA was heavily optimised to stay within IBM watsonx.ai's free Lite tier limits.

Metric	Before Optimisation	After Optimisation
Input tokens / call	~1,400	~505
History payload	Sent from browser (unbounded)	Held server-side (fixed window)
System prompt size	~650 tokens	~200 tokens
Truncation method	Character-count (inaccurate)	Word-count (token-accurate)
Output budget	280 tokens (fixed)	50 / 180 / 400 (dynamic per turn)
Full session cost	~18,000 tokens	~6,500 tokens
Sessions / month (free tier)	~2–3	~7–8

Key changes: server-side session store, compressed system prompt, word-count truncation, dynamic output caps per interview phase, background model warmup at startup.

---

📄 Resume Feature

ARIA supports optional resume upload to personalise the interview with questions drawn from your actual experience.

Step	What happens	Token cost
Upload PDF/DOCX/TXT	Text extracted server-side (PyMuPDF / python-docx)	0
Digest generation	ONE LLM call → ~120-word structured summary	~650 input + ~150 output (once)
Per interview call	Digest injected into system prompt	+~130 tokens/call
Full session w/ resume	~10,200 tokens total	~4–5 sessions/month

How it works:

1. Click 🚀 Start New Interview — creates your session
2. Upload your resume in the 📄 Resume Upload sidebar section
3. ARIA reads your Role, Experience, Skills, Projects, and Achievements from the digest
4. Questions are drawn from both general interview knowledge AND your specific resume
5. Final Report Card includes a Resume Alignment score — how well your answers matched your CV claims

Supported formats: PDF · DOCX · TXT (max ~5 pages recommended)

Resume is never stored — raw text is discarded after the digest is generated. Only the ~120-word summary is kept in memory for the session duration.

---

Built with ❤️ using Meta Llama 3.3 70B via IBM watsonx.ai

IBM SkillsBuild × AICTE Internship 2026