AI & Machine Learning.

RAG when answers depend on current, proprietary or frequently changing data (knowledge bases, policy documents, product catalogs). Fine-tuning when you need a model to adopt a tone, format, or specialized reasoning pattern that prompting can't reliably produce. Most enterprise problems are 80 percent RAG and 20 percent prompt engineering. Fine-tuning is expensive, locks you to a base model and is rarely the first move - we start with RAG and evaluate fine-tuning only against a measured baseline.

Hallucinations are mitigated, not eliminated. We use retrieval with strict source attribution, structured outputs validated against JSON schemas, confidence thresholds with deferral to humans, prompt-level guardrails (Guardrails AI, NeMo Guardrails) and an evaluation harness (Ragas, LangSmith, custom) that runs on every prompt change. For regulated workflows we add a human-in-the-loop checkpoint - if the cost of a wrong answer is high, the model should not be the last word.

We default to Anthropic Claude or OpenAI GPT for reasoning-heavy tasks, smaller fine-tuned open models (Llama 3, Mistral, Qwen) for cost-sensitive bulk classification, and embeddings from Cohere or text-embedding-3 for retrieval. For data-residency-sensitive clients we deploy open models on AWS Bedrock, Azure OpenAI in EU regions, or self-hosted on GPU instances. Model choice is per-workload, not per-vendor-loyalty.

Discovery and a proof-of-value are fixed-price (typically 4 to 8 weeks). Production builds are time-and-materials because the iteration loop on prompts, retrieval and evaluation is genuinely unpredictable. We also surface inference cost as a first-class budget line - a workflow that costs USD 50 per run in tokens may need re-architecting before launch, not after the bill arrives.

We address this contractually and architecturally. OpenAI, Anthropic and Google offer enterprise tiers with no training on customer data and EU data residency options. For NDPC, NDPA, POPIA or GDPR-regulated workloads we use Azure OpenAI in EU regions, Bedrock with VPC endpoints, or self-hosted open models. PII is redacted before any external API call when policy requires it, with full audit logs of what was sent.

Before shipping, we build a golden evaluation dataset with at least 100 representative examples and target metrics (accuracy, faithfulness, helpfulness, latency, cost per call). After shipping, we instrument user-level metrics (deflection rate for chat, task completion for agents, override rate for recommendations) and run weekly evaluation cycles. If you can't measure it, you can't operate it - and we won't ship it.

When deterministic logic, a search index, or a well-built form will do the job. AI is a poor fit for legally-binding decisions without human review, for low-volume workflows where the build cost dwarfs the savings, and for problems where you don't have evaluation data. We have walked clients back from LLM projects to a SQL query or a regex - that's good consulting, not lost revenue.