Skip to main content
Code Hippies

AI & LLM engineering

Retrieval-grounded AI features that answer from your data, not from imagination

RAG systems, AI assistants and LLM-backed workflows built so the model answers from retrieved source documents and says it does not know when nothing relevant is retrieved — with the evaluation harness to prove it.

Discuss this project Three to six weeks for a grounded assistant with an evaluation harness.

What you get

  • Retrieval pipeline: chunking strategy, embeddings, vector store, reranking
  • Grounded generation — answers cite retrieved chunks or decline
  • Refusal path when retrieval returns nothing relevant
  • Evaluation harness with a fixed question set, run in CI
  • Cost and latency instrumentation per request
  • Logged unanswered questions so content gaps become visible

Stack options

Chosen per project against your constraints. No single engagement uses all of these.

Models
Anthropic ClaudeOpenAIOpen-weight models via vLLM / Ollama
Retrieval
pgvectorSQLite + vector extensionManaged vector DBLocal lexical index
Orchestration
TypeScript / Next.js route handlersPython / FastAPI
Evaluation
Golden question sets in CIRetrieval hit-rate metricsHuman review queue

The chat widget on this site is the demonstration

The assistant in the corner of this page runs the same architecture I build for clients. It indexes a knowledge base of markdown documents — one per service, one per case study, one for the process and bio — retrieves the most relevant chunks for each question, and answers only from those chunks. Ask it something outside the knowledge base and it says it does not have that information and offers to connect you directly. That is the whole point.

Grounding reduces hallucination substantially — it does not eliminate it

Anyone selling you a guarantee of zero hallucination is selling you something they cannot deliver. What retrieval grounding does is make the failure mode honest: with no relevant source, the correct behaviour is a refusal rather than an invention, and that behaviour is testable. The evaluation harness runs a fixed question set on every deploy, including questions that should be refused, so a regression in the refusal path is caught by CI rather than by a customer.

Retrieval quality is the whole game

Most disappointing AI features are not model problems, they are retrieval problems: chunks split mid-sentence, no metadata, one embedding pass and no reranking. The work is unglamorous — chunking on semantic boundaries, keeping document titles attached to their chunks, measuring hit rate against a real question set — and it is where the improvement actually comes from.

Proof

AI & LLM engineering in production

Live builds where this work was actually done, with the technical signals verified on each site.

  • Health & Coaching2026

    FitWithNash — Consultation

    EngineeringAstro site (public source at github.com/codehippies11/fitwithnash) deployed to Vercel's edge network — the served page is essentially all content and almost no asset weight, and it carries the deepest structured-data graph in the portfolio: Person, ProfessionalService, Service, OfferCatalog, Offer, EducationalOccupationalCredential, Audience and a full FAQPage.

    In plain termsTurns a nutritionist's practice into a single page that answers every question a prospective client has, then books them.

    • Astro
    • Vercel
    • Static generation
    • JSON-LD / Schema.org
  • Marketing & Agency2026

    Core Media Solutions

    EngineeringContent-rich agency site (35 images, an enquiry form, a full section outline) whose entire commercial offer is mirrored in JSON-LD — Service, OfferCatalog, Offer, ContactPoint, PostalAddress, ImageObject and an FAQPage graph — so the service list is legible to search engines, not just to visitors.

    In plain termsExplains everything the agency sells, answers the usual questions up front, and captures enquiries on the same page.

    • Static site
    • Hostinger CDN
    • JSON-LD / Schema.org
    • FAQ schema
  • News & Publishing2025

    NewsLive24

    EngineeringWordPress origin behind nginx, instrumented with GA4 via Google Tag Manager and monetised with Google AdSense; the document carries the richest structured-data graph of the portfolio — NewsArticle, NewsMediaOrganization, CollectionPage, ItemList, ContactPoint and SearchAction — plus hreflang alternates.

    In plain termsRuns a full English news desk on a CMS the editors already know, while the technical SEO and ad plumbing is handled underneath them.

    • WordPress
    • nginx
    • GA4 / Tag Manager
    • Google AdSense

FAQ

Questions about ai & llm engineering

  • Will the AI make things up about my business?

    The system is built so it answers only from documents you control, and refuses when retrieval returns nothing relevant. That reduces fabrication substantially and makes the failure mode a visible refusal rather than a confident invention. It is not an absolute guarantee, and I will not claim it is — which is why the evaluation harness includes questions that must be refused.

  • Which model do you use?

    Whichever fits the constraint. The retrieval layer and the generation layer are kept separate, so the model is a swappable dependency rather than an architectural commitment.

  • What does it cost to run?

    Per-request token cost and latency are instrumented from the first build, so you see the running cost before launch rather than in the first invoice. Retrieval and caching are the two levers that actually move it.

Start a ai & llm engineering engagement

Send the brief with your project type, budget band and timeline. You'll get a scoped recommendation and a fixed price against a written scope.