All case studies
AI SaaS

AI Assistant ChatBot

Built an AI-powered chatbot SaaS product featuring a configurable knowledge base, embeddable widget for any website, and an intelligent caching system that achieves 90% token savings. Trained on 13+ years of development expertise for automated client interaction.

Results

Key innovations and optimizations achieved in AI chatbot development

Token usage reduction through caching
90%
Professional experience data trained
13 Years
Ephemeral cache duration per session
5 Minutes
Custom training interface for updates
Real-time
What I owned

Making Professional Experience Interactive

The Vision

Traditional portfolios and resumes are static and one-dimensional. I wanted to create an engaging way for potential clients and employers to understand my capabilities, ask specific questions about my experience, and get personalized insights about whether I'm the right fit for their projects.

The Innovation

By training an AI model on my entire 13-year professional journey - every project, skill, challenge, and achievement - I created an intelligent assistant that can have natural conversations about my experience. It's like having a personal representative available 24/7 to answer questions about my work.

Solving the Cost Challenge

Building an AI chatbot with OpenAI's API presented a significant challenge: token consumption costs. Every API call consumes tokens based on the amount of data processed, and constantly fetching context about my experience would be prohibitively expensive.

The breakthrough came with implementing cache control using ephemeral caching. When a user starts a conversation, the system caches the complete context data (my experience, projects, skills) for 5 minutes. During that window, subsequent messages reuse the cached data rather than reprocessing everything.

This single optimization reduced token consumption by up to 90%, making the chatbot economically sustainable while maintaining full conversational context. Users get fast, intelligent responses without the system burning through API credits on redundant data processing.

Optimizing Server Performance

The initial architecture had another problem: to answer questions about my experience, the system needed to fetch data from multiple API endpoints - projects from one source, skills from another, work history from a third. This created unnecessary server load and increased response time.

The solution was implementing vector-based data storage directly in Firestore. All relevant information about my professional journey is converted into vector embeddings and stored as a unified dataset. When the chatbot needs context, it retrieves everything from a single point with minimal server overhead.

To keep this data fresh, I built an auto-update mechanism that refreshes the vector data every hour. Additionally, there's a manual training interface where I can immediately update the model with new accomplishments, ensuring the chatbot always represents my most current experience.

Building the Training Interface

As my career progresses, the chatbot needs to learn about new projects and skills. Rather than requiring database updates or code changes, I built a dedicated training page that lets me feed new information to the model in real-time.

This training interface allows me to input daily work updates, new project milestones, acquired skills, and achievements. The system processes this information, updates the vector database, and immediately makes it available to the chatbot - keeping the AI assistant current and accurate.

This approach transforms the chatbot from a static representation to a living, evolving assistant that grows alongside my career. It's proactive rather than reactive, always reflecting my latest capabilities and experience.

What was built

ChatBot Capabilities

Advanced features making AI conversations intelligent and cost-effective

  • 13 Years of Experience

    Trained on complete professional journey spanning 13 years. Understands every project, technology, challenge overcome, and skill acquired throughout my career.

  • Natural Conversations

    Engages in human-like dialogue about technical expertise, project experience, and career achievements. Answers specific questions with contextual understanding.

  • Project Fit Analysis

    Intelligently assesses whether my skills and experience match project requirements. Provides honest recommendations about suitability for specific roles or projects.

  • 90% Token Savings

    Ephemeral cache control reduces token consumption by up to 90%. Caches conversation context for 5 minutes, eliminating redundant API processing costs.

  • Vector-Based Data

    Unified vector embeddings stored in Firestore provide single-point data retrieval. Reduces server load and improves response time compared to multiple API calls.

  • Auto-Update Mechanism

    Vector database automatically refreshes every hour, ensuring the chatbot always has current information without manual intervention.

  • Real-Time Training

    Dedicated training interface allows immediate model updates. Feed new accomplishments, projects, or skills and see them reflected in conversations instantly.

  • Continuous Learning

    Logs daily work and achievements through training interface. Creates a living, evolving AI assistant that grows more knowledgeable with my career progression.

How it was delivered

Development Journey

From concept to production-ready AI assistant with cost-effective optimization

  1. Phase 1: Core Development

    Week 1

    Built the foundational ExpressJS backend with OpenAI API integration and Firestore database setup. Established basic conversation flow and response generation.

    • ExpressJS server setup
    • OpenAI API integration
    • Firestore configuration
    • Basic chat interface
  2. Phase 2: Data Training

    Week 2

    Compiled and structured 13 years of professional experience into training data. Created comprehensive dataset covering projects, skills, achievements, and work history.

    • Experience data compilation
    • Training dataset creation
    • Initial model training
    • Context optimization
  3. Phase 3: Cache Implementation

    Week 3

    Implemented ephemeral cache control to dramatically reduce token consumption. Achieved 90% cost savings while maintaining conversation quality.

    • Cache control system
    • 5-minute session caching
    • Token usage optimization
    • Cost monitoring
  4. Phase 4: Vector Database

    Week 4

    Developed vector-based data storage system in Firestore. Consolidated multiple data sources into single-point retrieval for reduced server load.

    • Vector embeddings
    • Unified data storage
    • Auto-update mechanism
    • Performance optimization
  5. Phase 5: Training Interface

    Week 4

    Built admin training page for real-time model updates. Enabled continuous learning and immediate reflection of new professional achievements.

    • Training interface
    • Real-time updates
    • Manual refresh capability
    • Data validation
Hard problems

Technical Challenges & Solutions

Key obstacles overcome in building cost-effective AI chatbot

Token Consumption Costs

Challenge

OpenAI API charges based on token usage. Every message required sending the full context about my 13 years of experience, resulting in thousands of tokens per conversation. This made the chatbot economically unsustainable for a portfolio project.

Solution

Implemented ephemeral cache control that stores conversation context for 5 minutes. When a chat session starts, the system caches all background data once. Subsequent messages in that 5-minute window reuse the cached context instead of reprocessing everything. This single optimization reduced token consumption by 90%, making the chatbot viable while maintaining full conversational intelligence.

Server Load from Multiple Data Sources

Challenge

To provide comprehensive answers, the chatbot initially fetched data from multiple endpoints - projects API, skills API, testimonials API, work history API. Each conversation triggered numerous server requests, creating unnecessary load and slowing response times.

Solution

Designed and implemented vector-based data storage in Firestore. Converted all relevant professional information into vector embeddings and stored them as a unified dataset. Now the chatbot retrieves everything from a single point with one query, dramatically reducing server overhead and improving response speed while maintaining data richness.

Keeping AI Training Current

Challenge

As my career progresses, the chatbot quickly becomes outdated. New projects, skills, and achievements weren't reflected in conversations. Manually updating the database or retraining the model for each small change was impractical and time-consuming.

Solution

Built a custom training interface that lets me feed real-time updates to the model. I can input daily work accomplishments, new project milestones, or acquired skills through a simple form. The system processes this information, updates the vector database, and immediately makes it available to the chatbot - transforming it from a static snapshot to a continuously evolving AI assistant.

Data Freshness vs. Performance

Challenge

Needed to balance keeping the AI's knowledge current with system performance. Constantly updating the vector database would be expensive and potentially disruptive to active conversations. But waiting too long between updates would make the chatbot provide outdated information.

Solution

Implemented a hybrid approach: automatic hourly updates refresh the vector database with any changes from connected data sources, ensuring regular currency without overwhelming the system. For urgent updates like new project wins or major achievements, the manual training interface provides immediate refresh capability. This dual-update strategy maintains both freshness and performance.

Context Understanding Accuracy

Challenge

Training an AI on 13 years of diverse experience across multiple technologies, industries, and project types risked creating a model that provided generic or inaccurate responses. The chatbot needed to truly understand nuances in my experience to provide valuable insights.

Solution

Structured the training data with detailed context about each project, skill, and achievement. Rather than just listing technologies used, included information about challenges faced, solutions implemented, client relationships, and lessons learned. This rich context enables the AI to understand not just what I've done, but how and why - allowing for more accurate project fit assessments and meaningful conversations about my capabilities.

Stack

Technology Stack

Modern AI technologies optimized for performance and cost-efficiency

  • ExpressJS
  • Node.js
  • OpenAI API
  • Firestore
  • JavaScript
  • REST API
  • Vector Embeddings
  • Cache Control
  • Firebase Admin SDK
  • React
  • Next.js
  • Tailwind CSS
  • Git
Work With Me

Need a Technical Leader You Can Hold Responsible?

I'm a fractional CTO. I own the technical decisions and the code behind them: if you have a team, I lead it; if you don't, I am the team. Tell me where you are and what's stuck.

2 of 5 seats open

I take a maximum of five fractional clients at a time.