AI & Automation Integration
I integrate large language models into products that already have users — OpenAI, Anthropic Claude and Gemini APIs wired into content pipelines, support chat, document processing and automated internal workflows.
The interesting work is not the API call. It is everything around it: what happens when the model is slow, when it returns something malformed, when the same prompt costs three times as much next month, and how you keep a non-deterministic component from making your product unpredictable.
What can actually be automated with an LLM?
The tasks that work are the ones where a competent human could do the job from the text in front of them, the output gets checked before it matters, and being wrong occasionally is survivable.
That covers a lot: summarising and classifying incoming text, drafting content a person then edits, extracting structured fields from unstructured documents, routing support tickets, generating first-pass descriptions or replies at volume.
What does not work is anything requiring the answer to be correct every time with no human in the loop — pricing, compliance decisions, medical or legal judgements, arithmetic that matters. If you describe that kind of use case, I will tell you it is the wrong tool rather than build it and let you discover the failure in production.
How do you keep AI features from becoming expensive?
By treating tokens as a per-request cost line from day one instead of discovering it on the first real invoice.
In practice: caching responses for repeated inputs, routing easy requests to a smaller and cheaper model and reserving the frontier model for the hard ones, trimming context aggressively rather than sending whole documents by default, and putting per-user and per-account rate limits in before launch rather than after an incident.
I also instrument spend per feature, so when the bill moves you can see which feature moved it.
- Response caching on repeated or near-identical inputs
- Model tiering — cheap model first, escalate only when needed
- Streaming responses so the interface feels fast at the same cost
- Per-user and per-account rate limits, in place before launch
- Cost and latency tracked per feature, not just per month
How do you handle the model being unreliable?
Structured output with schema validation on the way back, so a malformed response is caught at the boundary and retried rather than written to your database.
Timeouts and fallbacks on every call, because provider outages are a normal operating condition, not an edge case. Where it matters, a second provider behind the same interface — the abstraction is cheap to build up front and painful to retrofit during an outage.
And a logged trace of prompt, response and cost for every call, so when a user reports a bad answer you can look at what actually happened instead of guessing.
Do you build RAG and document search?
Yes — retrieval over your own documents so the model answers from your content rather than its training data: ingestion, chunking, embeddings, a vector store, and the retrieval layer in front of the model.
The unglamorous half is what determines whether it works. Chunking strategy, keeping the index in sync as source documents change, handling permissions so a user cannot retrieve a document they should not see, and returning citations so a reader can verify the answer. Most disappointing RAG systems are retrieval failures, not model failures.
What you get
- AI feature integrated into your existing product and deployed
- Prompt and model configuration held in code, versioned and reviewable
- Structured output validation, retries and provider fallback
- Cost, latency and error monitoring per feature
- Rate limiting and abuse protection
- A written note on what the feature does badly, and where a human must stay in the loop
Tools I use for this
A good fit if
- An existing product that needs one AI feature built properly, not an AI strategy
- A manual internal process eating hours a week that could be drafted automatically
- Support or sales teams drowning in repetitive text triage
- A team with a working prototype that now needs cost control, error handling and monitoring before launch
Not a fit if
- Training or fine-tuning your own foundation model — that is an ML research engagement, not this
- Anything where a wrong answer has legal, financial or safety consequences and no human reviews it
- "Add AI to it" with no specific task in mind; I will ask what job it does before quoting, and if there isn't one I will say so
- Deepfake, impersonation or undisclosed synthetic-content work
Work that shows this
Catch-up
An automated SaaS platform to source clips, generate commentary in your voice, render an avatar, and publish.
Read the case studyMakanify CRM
A comprehensive real-estate CRM for builders and brokers: leads, clients, projects, tasks, documents and omnichannel communication in one place.
Read the case studyQuestions clients ask
Which AI provider should we use?
Usually whichever is strongest at your specific task, benchmarked on your own data rather than on a public leaderboard — the differences on real workloads rarely match the marketing. I build against a provider-agnostic interface regardless, so switching later is a configuration change rather than a rewrite. That matters more than the initial pick, because the ranking changes every few months.
Will our data be used to train their models?
Not on the standard API tiers of the major providers, which is different from their consumer chat products — that distinction catches people out. I will confirm the current terms for whichever provider we choose and put the retention and training position in writing before any of your data goes near it. If your data cannot leave your infrastructure at all, self-hosted open models are the route, and I will say so up front.
How much will the API cost us per month?
It is a function of requests per month, tokens per request and which model tier handles them. I estimate it during scoping using your actual expected volume and put a figure in front of you before we build, then instrument the real number so you can see it from launch day rather than at the first invoice.
Can you add AI to a product you did not build?
Yes — most of this work is on someone else's codebase. I read the code first, integrate at a boundary that does not entangle the new feature with your existing logic, and leave it removable. An AI feature that cannot be turned off independently is a liability.
Need ai & automation integrations?
Tell me what you're building and what's in your way. I reply within 24 hours with honest questions and a rough estimate — no sales sequence.