2020 — Where It All Started
In 2020, I was deep in Unity VR development when a project pushed me into AI territory for the first time — a VR Learning Studio. The goal was ambitious: give students inside a VR environment the ability to speak naturally, ask questions, and receive intelligent responses from a virtual AI tutor.
To build this, I had to stitch together several completely separate AI tools — each doing one narrow job. IBM Watson Speech SDK captured the student's spoken voice inside Unity and converted it to text in real time. Google AutoML was used to train custom text classification models — teaching the system to understand which subject or topic a student's question belonged to, so it could be routed correctly. Parallel Dots added a semantic analysis and emotion detection layer — understanding not just what the student said, but the sentiment and intent behind it. Google Dialogflow then matched the processed text to manually defined intents and returned a scripted response. Finally, Watson Text-to-Speech read the response back to the student as audio inside VR.
Five tools. Five API keys. Five billing accounts. Months of integration work — just to make one feature work reliably.
What I Was Actually Doing All Day
If you have never worked with Dialogflow, here is what "building conversational AI" actually looked like in 2020. For every single topic a student might ask about inside the VR classroom, I had to manually do all of this:
And after all that work? If a student said something even slightly different from my training phrases — the system would fail, return an empty fallback response, and break the entire immersive VR experience. It was a sophisticated decision tree wearing an AI costume. Powerful for its time, but brutal to maintain.
💡 The VR Learning Studio included voice-based Q&A, speech-to-text for student interactions, emotion-aware responses, and an AI virtual instructor — all running inside a Unity VR environment. Building it required Watson Speech SDK for voice capture, Google AutoML to train custom text classification models, Parallel Dots for semantic analysis and emotion detection, Dialogflow for intent matching, and Watson TTS to speak responses back. What sounds simple on paper took months to wire together and maintain reliably.
Late 2022 — The Moment Everything Changed
In November 2022, OpenAI released ChatGPT to the public. I remember the moment clearly. I typed a question into a browser — and it just answered. No intents. No training phrases. No fallback handling. No retraining. It understood context, remembered the conversation, and responded like a person.
The contrast with what I had spent months building in Dialogflow was almost hard to process. Everything I had manually configured — the intent trees, entity definitions, webhook handlers, training data — ChatGPT handled automatically, in real time, for any question, in any phrasing.
This was not an incremental improvement over Dialogflow. This was a completely different category of technology. Large Language Models had arrived — and the entire rule-based era I had been building in was suddenly obsolete.
2023–2024 — The Explosion
What followed was the fastest technology shift I have witnessed in 13 years of development. GPT-4 launched in March 2023 with dramatically improved reasoning. Anthropic released Claude — focused on safety and long-context understanding. Google launched Gemini with multimodal capabilities across text, images, audio, and video. Meta released open-source LLaMA models anyone could run locally. GitHub Copilot went mainstream inside every major IDE.
I started using OpenAI's API extensively during this period — integrating LLM-powered responses directly into applications. The difference from Dialogflow was immediate. No training data. No intent mapping. Just a well-written prompt, and intelligent, contextual output. Features that previously took me weeks to configure in Dialogflow could now be built in a few hours.
2024–2025 — AI Gets a Real Voice Inside VR
By 2024, a new kind of AI tool arrived that was built specifically for VR and game developers — Convai SDK. This was a significant moment for me personally, because it brought everything I had struggled to manually build back in 2020 into a single, purpose-built Unity SDK.
I integrated Convai into Unity VR projects for both NPC dialogue and real-time voice interaction. Characters inside VR could now hold full conversations with users — understanding what they said, responding intelligently in character, and maintaining context across an interaction. No intent trees. No Dialogflow. No Watson Speech pipeline. Just a natural conversation between a user and an AI-powered virtual character inside an immersive environment.
Compared to what I built in 2020 — where I spent months wiring five separate APIs together just to get a basic Q&A response — Convai felt almost effortless. The contrast was a direct, personal reminder of just how far AI had come in four years.
2026 — Where I Am Today
My AI workflow today looks nothing like 2020. I no longer use a single specialised tool for a single narrow task. Instead, I have a small stack of powerful models — each used for what it genuinely does best.
Claude is now my primary AI for building applications. I use it to create entire features, write and structure content at scale, reason through complex architecture decisions, and handle deep analytical work. Its long context window and ability to understand large codebases makes it genuinely useful for serious production work — not just quick answers. This entire site, AllInOneAICenter, is built with Claude as the core tool.
Google Gemini is my second opinion. When I want to cross-check an approach, validate a technical direction, or challenge an assumption before committing to it, I ask Gemini. Two different models thinking through the same problem surfaces gaps I might otherwise miss.
GitHub Copilot and Cursor handle real-time code checking, inline completions, and refactoring directly inside my IDE. What used to require a senior developer review for simple logic checks now takes seconds. For someone who spent months manually writing every Dialogflow webhook handler, this still feels remarkable.
Then vs Now — The Real Difference
- Manually define every intent in Dialogflow
- Write 10–15 training phrases per topic
- 5 separate tools for one voice feature
- 5 API keys and 5 billing accounts
- Parallel Dots for semantic analysis separately
- AutoML training required for text classification
- Broke when users phrased things differently
- Weeks of setup before first working demo
- No reasoning — pure pattern matching
- No intents needed — models understand naturally
- One prompt replaces hundreds of training phrases
- One model handles voice, text, vision, and code
- Single API, unified billing
- Semantic analysis built into the model
- Convai SDK for intelligent NPC voice in VR
- Understands any phrasing, any context
- Working prototype in hours, not weeks
- Genuine reasoning, memory, and context
The VR Learning Studio — Then vs Now
To make this concrete, here is exactly how my original VR Learning Studio project would be built today compared to 2020:
2020 — The Pipeline I Actually Built
Student speaks inside VR → Watson Speech SDK captures and transcribes audio inside Unity → AutoML classifies the text into the correct subject/topic → Parallel Dots analyses the semantic meaning and detects emotion/intent → text sent to Dialogflow → Dialogflow tries to match an intent from my manually written list → scripted response retrieved → Watson TTS converts response back to speech → audio plays in VR. Five systems. Weeks of setup. Constant failures when students asked anything unexpected.
2026 — How I Would Build It Today
Student speaks inside VR → OpenAI Whisper transcribes audio in real time → text sent directly to Claude or GPT-4o API with a system prompt defining the subject, the tutor's persona, and the learning context → intelligent, contextual, natural response returned in under a second → ElevenLabs or browser TTS speaks the response. Two systems. No intent training whatsoever. Students can ask absolutely anything — and get a genuinely helpful answer.
What Actually Changed — And Why It Matters
The shift from Dialogflow to modern LLMs is not just about better tooling. It is a fundamental change in how AI actually works. Dialogflow and Watson were rule-based systems — they matched patterns you defined. You wrote the rules. The AI executed them. If you did not write a rule for something, the system had absolutely no idea what to do.
Modern LLMs like Claude and GPT-4 are reasoning systems. They do not match patterns — they understand language, context, and intent at a deep level. They handle questions you never anticipated. They remember conversation history. They think through problems step by step. They can see images, read entire codebases, write and debug code, and generate structured content. They get smarter over time without you doing anything.
For someone who spent months building a VR voice bot by manually wiring Watson Speech SDK through Dialogflow to Watson TTS — what is available in 2026 is almost hard to believe. What took a development team months of setup in 2020 can genuinely be built by one developer in a single afternoon.
What Has Not Changed
Here is something worth saying clearly: the thinking skills from 2020 still matter. Understanding how to design a conversation. Knowing what a user actually needs versus what they literally say. Thinking carefully about context, memory, and edge cases. Knowing when AI is right and when it is confidently wrong. Building AI into a real product that actual users depend on.
The difference is that in 2020, I spent 80% of my time fighting the tools and 20% solving the real problem. Today that ratio has completely flipped. The tools are powerful enough that I spend almost all of my time on the actual user experience — which is exactly where a developer's time should go.
Frequently Asked Questions
What was Dialogflow used for in 2020–2021?
Dialogflow was Google's NLP platform for building chatbots and voice assistants. Developers manually defined "intents" — categories of what users might say — along with training phrases and scripted responses. It worked within those limits but was rigid and required enormous manual effort for even moderately complex conversations.
What was IBM Watson Speech SDK used for?
Watson Speech SDK was IBM's tool for converting spoken audio into text (Speech-to-Text) and converting text back into spoken audio (Text-to-Speech). It was commonly paired with Dialogflow or Watson Assistant — you needed Watson for the voice layer and a separate NLP tool for understanding the meaning of the transcribed text.
What is AutoML and how was it used in VR projects?
Google AutoML allowed developers to train custom text classification models on their own labelled data without deep ML expertise. In the VR Learning Studio, it was used to classify a student's question into the correct subject area before routing it through the Dialogflow intent matching pipeline.
What is Parallel Dots and what was it used for?
Parallel Dots is an NLP API platform that provides semantic analysis, emotion detection, sentiment analysis, keyword extraction, and intent classification. In the VR Learning Studio, it was used to understand the deeper meaning and emotional tone behind a student's spoken question — going beyond simple keyword matching to understand what the student actually needed.
What is Convai SDK and how does it work in Unity VR?
Convai is an AI platform purpose-built for game and VR developers that enables intelligent, voice-driven NPC characters inside Unity. Instead of scripting rigid dialogue trees, developers give a character a persona and knowledge base — and Convai handles real-time voice input, natural language understanding, and contextual responses. It replaces the entire Watson + Dialogflow pipeline from 2020 with a single SDK integration.
How is Claude different from Dialogflow?
Dialogflow requires you to manually define every possible conversation path before deployment. Claude understands language naturally — you provide a system prompt describing the context and purpose, and it handles any conversation intelligently from there. No intent training, no fallback mapping, no retraining cycles. It also genuinely reasons rather than just pattern-matching.
Is Dialogflow still worth learning in 2026?
Dialogflow CX still has a place in strict enterprise environments — call centres, regulated industries — where conversation paths must be fully controlled and auditable. But for most conversational AI use cases, modern LLMs produce dramatically better results with a fraction of the setup time. New developers should start with LLM APIs, not intent-based platforms.
What AI tools should a Unity / VR developer use in 2026?
For conversational NPCs and voice interactions inside VR: Claude or GPT-4o API combined with OpenAI Whisper for speech recognition. For coding assistance inside Unity: GitHub Copilot or Cursor. For shader and C# script generation: Claude or ChatGPT with a Unity-specific system prompt. For 3D asset generation: Meshy or Luma AI.