# AI, engineered for production

> We build AI products, integrate AI into existing systems, and know the engineering it takes to make them trustworthy — grounded answers, fail-soft design, controlled costs.

## Build, integrate, advise

**Build.** AI products end-to-end: RAG answer engines with cited sources, semantic search, and AI content pipelines — from feasibility to live operation.

**Integrate.** Add AI to the systems you already run — LLM-powered features, semantic search over your own content, automation pipelines. No rewrite required.

**Advise.** Honest guidance on where AI fits your product, what it will cost to run, and where a simpler solution beats a model. Strategy grounded in systems we have shipped.

## How production RAG works

1. **Understand the question.** A language model rewrites the question into search terms and extracts anything structured — cached, so repeats cost nothing.
2. **Retrieve from a curated index.** The query runs against a continuously updated, trust-scored index — fast, cheap and deterministic.
3. **Generate only from the sources.** The model composes a concise answer using nothing but the retrieved documents. No sources, no answer — never a guess.
4. **Cite everything, stream it live.** Inline citations link every claim to its source while the answer streams token by token.

Long-form explanation: https://devnest.ro/insights/how-production-rag-works

## Capabilities running in production today

Not a capability list — a production inventory. Each of these runs live.

- **Grounded RAG answers** (retrieval-augmented generation) — plain-English questions answered only from a curated, trust-scored index, cited inline and streamed in real time.
- **Interpretable semantic matching** (embeddings) — content embeddings projected onto a curated vocabulary of business concepts, so every match comes with a human-readable explanation.
- **Automatic transcription** (Whisper) — podcasts and video transcribed automatically, so spoken content becomes searchable text feeding the search index and AI summaries.
- **AI narration and timed captions** (text-to-speech) — voiceovers generated from a script, then captioned word-by-word against the rendered audio, with fallbacks so captions never fail.
- **Automated social video** (media pipeline) — a single article becomes a rendered social reel: AI-written script, narration, synced captions and assembly with ffmpeg.
- **Tagging, summaries and abstracts** (content enrichment) — articles and videos summarised, keyword-tagged and described automatically at scale, with editorial edits always preserved.
- **Semantic audience matching** (Elasticsearch kNN) — new content matched to the communities most likely to care, ranked by vector similarity with named shared topics for explainability.
- **AI-assisted newsletters** (automation) — newsletters assembled automatically with AI-written subject lines and openers, reviewed by people, shipped on schedule.

## Engineering principles

**Grounded, never free-form.** Where AI writes anything a user reads, it is constrained to supplied source text. Facts come from your data; the model arranges them.

**Fail-soft by design.** Every AI call has a non-AI fallback. A provider outage degrades quality — plainer copy, keyword search — but never breaks the product.

**Cost under control.** Caching, heuristic gates, per-user rate limits and model tiering keep spend predictable and proportional to the value delivered.

## AI FAQ

**What is retrieval-augmented generation (RAG)?** An architecture where an AI model answers only from documents retrieved from a curated index, rather than from its own memory. Every answer is grounded in real sources and cites them, keeping responses accurate, current and verifiable. We used RAG to build the Business111.com answer engine.

**How do you stop AI hallucinating?** By never letting the model answer free-form. We constrain generation to retrieved, trust-scored documents, require inline citations, and return a safe default when the sources cannot support an answer. Grounding plus citations means every claim is traceable to a source a human can check.

**Can you add AI to our existing product?** Yes — and without a rewrite. We typically start with one high-value integration: semantic search over your content, an assistant grounded in your documentation, or an enrichment pipeline for your data. AI components run alongside your existing stack behind clear interfaces, engineered fail-soft. First integrations typically reach production in weeks, not quarters.

**Which AI models and providers do you work with?** OpenAI and Anthropic models for generation, OpenAI embeddings for semantic search, Whisper for speech-to-text, and Elasticsearch for lexical and vector (kNN) retrieval. We choose per use case and keep model choices configurable, so providers can be swapped as the market moves.

**What does a typical AI project involve?** A short feasibility phase first: where AI genuinely helps, what data you have, and what it will cost to run. Then an incremental build — retrieval and grounding first, generation second, evaluation throughout — delivering working software at each stage rather than a demo at the end.

**How do you keep AI running costs predictable?** Caching, so repeat questions never hit the model twice; heuristic gates that skip AI calls when a simpler check can decide; per-user rate limits; and model tiering — small models for routine steps, larger ones only where quality is user-visible.

**What happens when the AI provider has an outage?** Nothing breaks. Every AI call has a non-AI fallback: search falls back to keyword results, generated copy falls back to neutral templates, captions fall back to timed text. Quality degrades gracefully; features never hard-fail.

**Do we own the system you build?** Yes. Full ownership transfers to you: code, infrastructure, prompts, evaluation suites and operational runbooks. No managed AI vendor sits between you and your features, and there is no lock-in to us either.