Building an AI Agent for My Portfolio — Engineering Log

Most portfolio websites are static. Recruiters visit, read, and leave.

I wanted something different:

A real-time AI agent that answers recruiter questions about me, my projects, and my skills — even when I’m offline.

So I built a lightweight production-style AI assistant using Groq API, Vercel serverless functions, and Telegram notifications.

System Architecture

Recruiter opens kshitij.info

↓

Clicks floating chat widget

↓

Message sent to /api/chat (Vercel Serverless Function)

↓

Server injects system prompt + conversation history

↓

Groq LLaMA-3.1 model generates response

↓

Response returned to widget

↓

Conversation transcript sent to Telegram

Tech Stack

Goal:

Build a recruiter-aware conversational assistant with zero infrastructure cost.

Designing the System Prompt

The system prompt defines how the assistant behaves.

Instead of acting like a generic chatbot, the agent responds as me.

Example identity block:

You are Kshitij Mishra.

You bridge the gap between AI systems and real-world reliability.

Rules implemented:

Salary questions → redirect to email

Availability questions → suggest contact directly

Irrelevant questions → polite redirection

Responses always end with a follow-up question

This ensures structured recruiter conversations.

Backend Implementation

Serverless endpoint:

/api/chat

Responsibilities:

Receive messages from widget

Send request to Groq inference API

Return generated response

Send transcript to Telegram

Example request flow:

messages = [
system_prompt,
conversation_history,
latest_user_message
]

Stateless architecture keeps deployment simple.

No database required.

Groq API Integration

Inference endpoint:

https://api.groq.com/openai/v1/chat/completions

Model used:

llama-3.1-8b-instant

Reasons:

Extremely fast response time

OpenAI-compatible API format

Free-tier friendly

Reliable multi-turn conversations

Chat Widget Engineering

The widget is written using pure JavaScript.

Features:

Floating assistant bubble

Dark theme UI matching portfolio

Conversation memory stored locally

Quick recruiter question buttons

Auto tooltip prompt after page load

Transcript trigger on page exit

Conversation state example:

messages[]

Sent with every request.

Telegram Notification Pipeline

Telegram acts as a lightweight observability layer.

Setup steps:

Create bot using @BotFather

Retrieve chat ID from @userinfobot

Store credentials in Vercel environment variables

Example notification:

Recruiter: Tell me about yourself

Agent: I'm Kshitij Mishra...

This allows real-time visibility into recruiter interactions.

Deployment Workflow

Backend:

GitHub repo → imported into Vercel

Auto-deploy on commit

Frontend:

Widget added inside portfolio index.md

Deployment time:

~2 minutes

Environment Variables

Stored inside Vercel dashboard:

GROQ_API_KEY

TELEGRAM_BOT_TOKEN

TELEGRAM_CHAT_ID

No secrets exposed inside repository.

Challenges Faced During Development

API Key Exposure Risk

Initial debugging revealed keys should never appear inside frontend code.

Resolved using:

process.env.*

inside serverless runtime.

Model Deprecation Issue

Deprecated model:

llama3-8b-8192

Migrated to:

llama-3.1-8b-instant

Improved speed and reliability.

Rate Limits on Gemini API

Gemini free tier caused interruptions during testing.

Switched inference provider to:

Groq API

Result:

Stable recruiter session handling.

Debugging Serverless Runtime Issues

Primary debugging tools used:

Vercel Function Logs

Browser Network Inspector

Direct endpoint testing:

https://kshitij-ai-agent.vercel.app/api/chat

Used to validate:

payload structure

environment variables

response latency

error handling

Performance Observations

Latency:

~400–900ms

Cold start:

Minimal

Cost:

$0 per month

Scaling:

Automatic

Future Improvements Planned

Session analytics dashboard

Vector-based memory retrieval

Resume embedding search

Conversation scoring signals

Multi-agent routing

Final Result

Live AI recruiter assistant running on:

kshitij.info

Capabilities:

Real-time responses

Portfolio-aware answers

Automatic transcript delivery

Zero infrastructure cost

Sub-second latency