Ayush Sharma · Product & Operations Lead, Medoc Health IT
A running workbook, not a polished essay — this is the actual shape of my AI rabbit hole, from LinkedIn commentary to RAG pipelines. I'll keep adding to it as I learn more.
I started posting about AI and LLMs on LinkedIn back when Google's chatbot was still called Bard, not Gemini, and GPT was still early enough in its lifecycle that most people outside tech hadn't touched it yet. I made a handful of predictions in that window about where this was heading — some of them landed, some didn't, and the ones that landed taught me more than the ones I got right for the wrong reasons.
You all have been using ChatGPT wrongly!: https://www.linkedin.com/posts/ayush-sharman_ai-chatgpt-work-activity-7057971346760245248-1A2i
Google Bard is better than ChatGPT but I still use ChatGPT: https://www.linkedin.com/posts/ayush-sharman_ai-chatgpt-bard-activity-7083043513822568448-rwCX
What changed over time wasn't the interest — it was the depth. Early posts were commentary: reacting to a launch, a benchmark, a demo. Somewhere in that stretch, commentary turned into tinkering — actually building with the tools instead of just writing about them. That's the shift this document tries to capture: Part 1 stays close to where I started (concepts, terminology, which tool actually does what), and Part 2 goes into where I've ended up spending most of my time lately — RAG, agent orchestration, and the plumbing underneath the products I use every day.
One influence worth naming directly: a mentor from my time at Clear posts regularly about practical AI tooling — not the hype layer, the "here's exactly how you wire this up" layer. One of his walkthroughs, on bridging WhatsApp into Claude, is the reason Part 2 has an actual worked example instead of just theory. Credited and linked in Part 2.4.
Every field builds a wall of jargon that's mostly simple ideas wearing a costume. Here's the vocabulary I actually use, stripped down.
| Term | Plain meaning |
|---|---|
| LLM (Large Language Model) | A model trained on huge amounts of text to predict what comes next. Everything else in this table is either a way of controlling that prediction or a way of connecting it to the outside world. |
| Token | The chunk of text a model actually processes — roughly ¾ of a word in English. Cost, speed, and context limits are usually measured in tokens, not words. |
| Context window | How much text (in tokens) a model can "see" at once — the prompt, the conversation history, any documents attached. Once you exceed it, the oldest content falls out of view. |
| Prompt vs. system prompt | The system prompt sets standing behavior before the conversation starts (tone, role, rules). The user prompt is the specific ask in the moment. Confusing the two is why some instructions "don't stick." |
| Temperature | A dial for randomness in the model's output. Low temperature = more predictable, repeatable answers. High = more varied, more creative, less consistent. |
| Fine-tuning | Further training a model on a narrower dataset so it specializes in a task or tone. Increasingly less common for product work now that in-context methods (below) do a lot of the same job more cheaply. |
| Embeddings | Turning text into a list of numbers (a vector) that captures its meaning, so "semantically similar" text ends up numerically close. This is the math underneath search and RAG. |
| RAG (Retrieval-Augmented Generation) | Instead of relying only on what a model memorized during training, you retrieve relevant documents at query time and hand them to the model as context. Covered in depth in Part 2.1. |
| Hallucination | A model stating something false with the same confidence as something true. Not a bug you patch once — a standing property of how these models generate text, which is why guardrails and retrieval exist. |
| Agent | A model that doesn't just answer — it takes actions (calling tools, running code, querying a database) and decides what to do next based on the result, in a loop. Covered in Part 2.2. |
| Orchestration | The layer that manages multiple steps, tools, or agents working together — deciding sequencing, retries, and handoffs, so the "thinking" and the "coordinating" aren't the same piece of code. |
| MCP (Model Context Protocol) | An open standard, introduced by Anthropic in late 2024, for connecting an AI model to external tools and data sources through one consistent interface instead of a custom integration per tool. Covered in Part 2.3. |
| Multimodal | A model that can take in or produce more than just text — images, audio, video. |
| Chain-of-thought | Prompting or training a model to reason through intermediate steps before answering, rather than jumping straight to a conclusion. Improves accuracy on anything that benefits from working shown. |
| Few-shot prompting | Giving the model a small number of worked examples in the prompt itself, instead of relying purely on instructions. Often more reliable than describing what you want in the abstract. |
| RLHF (Reinforcement Learning from Human Feedback) | A training step where human preferences (which of two answers is better) are used to steer the model toward more helpful, less harmful outputs. Part of why modern models feel less robotic than raw next-token predictors. |
| Guardrails | Explicit constraints — in the prompt, the system, or a separate checking layer — that stop a model from doing something it technically could do but shouldn't in this product context. |
| Vector database | A database built to store embeddings and search by semantic similarity instead of exact keyword match. The retrieval half of RAG runs on top of one. |
Nobody needs all of these — but knowing what each category is actually good for stops you from using a hammer on a screw. This reflects the landscape as of mid-2026; this space moves fast enough that it's worth re-checking every few months rather than trusting a static list.
| Use case | What I reach for | Why |
|---|---|---|
| General assistant / writing / analysis | Claude, ChatGPT | Claude for long-form writing, document analysis, and anything where reasoning quality on a messy real-world input matters more than speed. ChatGPT for breadth and multimodal range (text, image, voice) in one place. |
| Coding | Claude Code, Cursor | Claude Code for repository-aware, multi-file, terminal-native work. Cursor for an AI-native IDE experience when I want the editor itself to be AI-first. |
| Real-time research / fact-checking | Perplexity | Built around live search with citations by default, which matters when the answer needs to be checked, not just plausible. |
| Document Q&A over my own material | NotebookLM | Answers strictly from a set of documents you feed it, rather than blending in outside training knowledge — useful specifically when you want the answer grounded in your material, not the model's general memory. |
| Automation / connecting tools without writing a backend | n8n | A visual workflow builder that's picked up AI-agent nodes and native MCP support, so it sits well for automations that are one step inside a bigger business process rather than a standalone agent. |
| Image generation | Midjourney, GPT Image | Midjourney for stylized, art-directed output. GPT/DALL·E-family tools for fast iteration and text-in-image accuracy. |
| Voice / TTS | ElevenLabs | The default for realistic voice generation and voice agents at this point. |
| Knowledge work agent (multi-step tasks across tools) | Claude Cowork | For handing off heavier, multi-step tasks — research, drafting, coordinating across files — rather than a single back-and-forth chat. |
A pattern worth naming: the winning setup for most people isn't one tool that does everything, it's one primary assistant plus one or two specialists, with a clear point where you hand off from one to the other. Trying to force a single tool to cover writing, coding, image work, and automation usually means you're paying for capability you don't use in three of the four categories.