AI that does a real job inside a real product
I know AI inside out, and I build products both with it and around it: algorithms that find patterns in data, RAG applications that answer from your own content, models trained and fine-tuned for a specialist task, and analysis that turns scattered data into decisions. I bring the product side (users, data, evaluation, cost) and the hands to build the first version myself.
What I build
Seven kinds of AI work, each one shaped around the people who will use it.
-
RAG applications
Assistants and search that answer from your own documents, help articles, policies or product data, and show where each answer came from. Retrieval-augmented generation keeps answers grounded in your content, and current whenever that content changes.
Good for Support assistants, internal knowledge search, product Q&A, onboarding guides.
-
Trained and fine-tuned models
When a general model isn’t the right fit, I build the training data, fine-tune a smaller specialised model, review what it produces and iterate. A focused model can be more consistent, and much cheaper to run at scale.
Good for Specialist language, consistent formats, high-volume tasks where cost matters.
Where I’ve done it: AI financial analysis
-
Pattern recognition and algorithm design
Algorithms that find patterns and signals in user behaviour, search, sales, market and financial data, and turn them into product decisions and features. The point is never the algorithm itself, but what it lets people see sooner.
Good for Spotting trends early, flagging anomalies, ranking and matching, deciding what to build next.
Where I’ve done it: AI financial analysis and horsesales.ie
-
AI-powered analysis
Pipelines that pull many data sources together and turn them into clear insight: financial data for consulting clients, and user behaviour, search terms, listings and sales data for my own marketplace.
Good for Dashboards that explain themselves, market monitoring, product decisions from data.
Where I’ve done it: AI financial analysis and horsesales.ie
-
AI features inside existing products
Summaries, smart search, classification, recommendations and drafting tools, placed where users already work. AI should remove a real step from someone’s day, not sit in a separate chat window nobody opens.
Good for Marketplaces, content platforms, internal tools, customer portals.
-
AI promotional videos
Marketing videos for web products, made end to end with AI. I write the script and produce the whole video myself, from the visuals to the sound and the final edit, using generative video, image and audio tools on my own GPU workstation. I’ve made promotional videos for clients who sell online courses, and for my own products.
Good for Promoting online courses and web apps, product launches, social media ads, explainers.
Where I’ve done it: Earlier projects
-
Products AI assistants can find
Structured content and data so a product is found, cited and recommended by Google AI Overviews, ChatGPT and Perplexity. This website is built the same way.
Good for Any product that needs traffic in a world where people ask AI first.
Where I’ve done it: horsesales.ie
RAG, fine-tuning or just a good prompt?
The most useful AI decision is often the least exciting one. Pick the need that sounds most like yours, and here’s where I’d start.
Where I’d start
RAG
Retrieval finds the right passages at the moment someone asks, so answers stay current and can cite their source. When a document changes, nothing needs retraining.
Where I’d start
Fine-tuning
Training a model on good examples builds the behaviour in. A smaller fine-tuned model is often faster and cheaper to run than a large general one.
Where I’d start
Prompting, with a test set
Start here. Well-designed prompts, checked against real questions, get you to launch fastest and show whether anything heavier is worth paying for.
Where I’d start
An AI analysis pipeline
Bring the sources together and clean them, then let AI surface patterns and anomalies for people to act on. This is the kind of product I built for financial-data clients.
From idea to launch
AI products fail in familiar ways: no clear job, poor data, no way to tell whether answers are good. This is the order I work in to avoid that.
-
Start from the job
What task should AI take off someone’s plate, and how will we know it did it well? Some ideas don’t need AI at all, and finding that out early is a good result too.
-
Look at the data
An AI product is only as good as the content behind it. I check what exists, how clean it is, who owns it and what has to stay private.
-
Pick the simplest approach that works
A well-designed prompt first. RAG when answers must come from your own content. Fine-tuning when you need specialist behaviour or lower running costs.
-
Build the test set before the product
A set of real questions with known good answers, so every change to prompts, data or model is measured instead of guessed.
-
Design for trust
Show sources, handle “I don’t know” gracefully, keep a person in the loop where mistakes are costly, and protect personal data under GDPR.
-
Ship small, measure, improve
Track answer quality, user satisfaction, response time and cost per query, then improve one measured step at a time.
What I work with
The tools change every few months. The judgement about when to use them doesn’t.
- Claude and other LLM APIs
- Embeddings and vector search
- RAG pipelines
- Model fine-tuning
- Local LLMs
- AWS AI services
- Python
- Node.js
- ComfyUI
- AI video generation
- Claude Code
- Cursor
- Copilot
AI questions people ask me
Straight answers, for hiring managers and for anyone deciding whether their product needs AI.
Can Ryan Mahato build AI products?
Yes. Ryan manages and builds AI products end to end: RAG applications that answer from a company’s own content, trained and fine-tuned specialised models, AI-powered data analysis, and AI features inside existing products. He handles the product side (users, data, evaluation, cost) and can build the first version himself.
Has Ryan trained or fine-tuned AI models?
Yes. As a consultant he fine-tuned a specialised, compute-efficient model on a corpus he helped build, reviewed its outputs and iterated, as part of AI analysis products for financial data. He also works with local LLMs and AWS AI services.
What is a RAG application, and when does a product need one?
RAG (retrieval-augmented generation) lets an AI model answer from your own documents by looking up the relevant passages at the moment someone asks. It suits support assistants and knowledge search, where answers must be current and traceable to a source.
How does Ryan manage an AI product?
He starts from the job the AI should do, checks the data, picks the simplest approach that works (prompting, RAG or fine-tuning), builds a test set before the product, designs for trust and privacy, and measures answer quality, user satisfaction, response time and cost per query.
Can Ryan make AI videos to market a product?
Yes. Ryan writes the script and produces promotional videos end to end with generative AI, from the visuals to the sound and the final edit. He has made them for clients who sell online courses, to market their courses and the web platforms they run on, and for his own products.
Does Ryan use AI in his own work?
Every day. He uses Claude Code, Cursor and Copilot to prototype, research, analyse data and write, which lets him test an idea in an afternoon.
Have an AI idea, or an AI product that isn’t landing?
I can help you decide what to build, build the first version and measure whether it works, as a full-time hire, a part-time product lead, a contractor or on a freelance project.