AI.
Two offerings for the "demo works, production breaks" gap. AI / ML engineering owns the model layer - serving, RAG, evaluation, cost, guardrails. Agentic systems owns the layer above - orchestration, tool use, planning, human-in-loop. Together they're how AI stops being a slide.
AI / ML engineering
Stack we live in
Anthropic API · OpenAI · Bedrock · Sagemaker · vLLM · Ray · LangChain · LangSmith · Weights & Biases · Pinecone · Qdrant · Weaviate
- Model serving + versioning at request-per-second scale
- RAG pipelines - indexing, retrieval, chunking, re-ranking
- Evaluation harness that measures what actually matters to the product
- Guardrails + safety - prompt-injection defence, PII redaction, output validation
- LLM cost architecture - caching, routing, tier-down for cheap tasks
- Fine-tuning + adapters when they're actually justified (usually they're not)
The demo works, production breaks. Or the LLM bill is climbing without a matching quality signal. Or the model's picked but nobody agrees what "good" looks like.
Agentic systems
Stack we live in
Claude Agent SDK · LangGraph · CrewAI · Temporal · OpenAI Assistants · MCP servers · Anthropic tool use · Zep memory
- Tool use design - what agents can call, when, with what confidence gates
- Orchestration (state graphs, event loops, task queues)
- Multi-step planning with graceful failure + retry
- Human-in-loop checkpoints - approval, escalation, override
- Agent observability - trace every tool call, every decision, every retry
- Evaluation of agent runs (not just single responses)
A workflow needs multi-step reasoning that can't be scripted. Or the humans-in-the-middle are the bottleneck and you want them reviewing, not typing.
Model layer, or agent layer?
Tell us where production keeps breaking - the model, the retrieval, the guardrails, or the orchestration. A principal architect replies within one working day.
Book a call