AI Engineering in practice
TECHNICAL CAPABILITY MAP
Reliable AI products are systems, not prompts.
A recruiter can see the engineering surface; a tech lead can follow each claim to the real architecture, controls, and known gaps.
Prompt Engineering — make the model's job explicit
Prompt Engineering is the discipline of defining instructions, context boundaries, output contracts, and versioning so an LLM behaves predictably. Here it means versioned prompt registries, not ad-hoc strings in application code.
Where and how: PromptDesk is the reference: its chat worker takes a registered prompt, fences untrusted customer context, limits the model to text, and validates results before use. See guardrails and PromptDesk architecture.
VERIFIED Versioned prompts + fenced context. NOT FOUND autonomous tool use.
Guardrails — constrain inputs, outputs, and spend
Guardrails are deterministic and operational controls around an LLM: screen unsafe input before a call, constrain the prompt, parse output strictly, fail closed, and cap requests. They are not a claim that the model is inherently safe.
Where and how: the shared baseline is Guardrails; Argus and Quizzeira align to its contracts. Model ladder & budgets shows provider order, timeouts, and request budgets.
Eval — decide when generated content can move forward
An eval is a measurable acceptance step, separate from generation. It makes quality visible and prevents a draft from silently becoming a user-facing result.
Where and how: Quizzeira generates a draft bank, then applies an eval gate before publishing study content. Its verified thresholds are documented with the pipeline rather than presented as generic AI claims.
RAG, chunks, and vector databases — retrieve only when evidence supports it
RAG augments a model with retrieved material. A chunk is a retrievable slice of source content. A vector database stores embeddings so semantically similar chunks can be found. These three terms are related, but having embeddings does not prove a live RAG path.
Where and how: Quizzeira has 768-dimensional embeddings, but retrieval is [UNUSED] on its generation path. Argus has partial RAG plumbing (RAG_LIMIT=5) but no query embedding on the live path. PromptDesk has no vector store. See comparison matrix and not in scope.
PARTIAL Argus retrieval. UNUSED Quizzeira generation search. NOT FOUND PromptDesk RAG.
What this demonstrates
The portfolio is intentionally evidence-backed: model orchestration, cost/rate controls, evaluation gates, security boundaries, observability, and human review are documented as engineering systems. Where a capability is missing, the documentation says so.