AI features that run in production, not just in the demo.
Semantic search, recommendations and assistants that answer from your own content, built on your stack with evals, cost control and a plan for when the model is wrong. I built and operate AI Search, used by stores worldwide.
- Free 30-min call
- No obligation
- Reply within 1 business day
A good fit if
Sound familiar?
“You have an AI demo that works on ten documents and falls over at ten thousand.”
“Your store’s search can’t handle how people actually describe products.”
“Support or sales wants an assistant that answers from your docs, and legal wants to know where the data goes.”
“Someone pasted an API key into a plugin and the bill nobody expected arrived.”
What I do
The work
Semantic and hybrid search
Embeddings plus keyword search, ranked so existing results are never lost. pgvector, Pinecone or a managed index, chosen for your catalog size and budget.
Assistants that cite
Retrieval-augmented assistants that answer from your content, show their sources, and say “I don’t know” instead of inventing. Guardrails, logging and feedback built in.
Recommendations and content discovery
Related products and articles by meaning rather than category, with rules you control for stock, margin and seasonality.
Evals and cost control
A test set of real queries before anything ships, latency and cost budgets per request, caching and batching so the feature stays affordable at scale.
Data boundaries
What leaves your infrastructure, to whom, and why, written down before the first API call. Secrets in your vault, least-privilege keys, per-site quotas.
How it works
From first call to shipped work
Feasibility, one week
Your data, your queries, a quick prototype on a sample, and an honest answer about whether AI is the right tool. Sometimes it isn’t.
Build with evals
The feature built in your repo, measured against the test set every week, with cost per query tracked from day one.
Operate
Monitoring, fallbacks and a runbook so a provider outage degrades gracefully instead of taking the site down.
Search features ship in three to six weeks; assistants in six to ten. Fixed price for defined work, hourly for ongoing engineering. How I work with US teams.
Proof
“Sam is an excellent programmer who I would have no hesitation recommending and look forward to working with again. He really thinks things through carefully and delivers great work.”
Client via Codeable, 5/5
Questions
Questions I get on the first call
Which models and providers do you use?
OpenAI and Anthropic APIs most often, open models where data has to stay in-house. The choice is made per feature on quality, cost and where the data is allowed to go.
How do you keep costs predictable?
Caching, batching, small models where they are enough, hard per-request budgets and usage dashboards. The AI Search service runs on a free tier of 10,000 items per site because the costs are engineered, not hoped for.
What about our customers’ data?
We agree the data boundary first: what is embedded, what is sent to a provider, what is stored and for how long. Nothing leaves your infrastructure without a written decision.
What does it cost?
Feasibility is a small fixed fee. Features are fixed price after that. Running costs are estimated up front and measured in production.
Tell me about the project.
Thirty minutes, free, no pitch. You’ll leave knowing what it would take.
Or email hello@samuelsilva.pt