ADVANCED LLM TRAINING
Advanced large language model engineering
We help experienced LLM practitioners compare advanced retrieval, fine-tuning and agent approaches through practical experiments and evaluation.
- A framework to diagnose the right approach before you build
- Deep, hands-on modules on fine-tuning, retrieval, and agents
- Learn-while-building on a real problem from your roadmap
When to go deep, and when not to.
Most quality problems have a cheapest-first answer. We teach practitioners to diagnose before they build, because the wrong fix is expensive and often doesn't help.
The rule we drill: exhaust prompting and retrieval before tuning, and don't build an agent for a task a chain handles. Complexity is a cost, not a badge. This program assumes you have shipped before; if you haven't, start with Building LLM Applications.
Diagnose before you build.Complexity is a cost, not a badge.
Once the decision is made, go deep on the technique that fits.
Hands-on, on real problems, on the tools you will use in production.
Practitioners work on tools they will use in production: LangChain and LangGraph, vector stores such as pgvector or Pinecone, tuning frameworks, and LangSmith or Langfuse for evaluation, adapted to your stack. These run best on production-grade infrastructure from Scaled GenAI and AI Platforms.
Fine-tuning
LoRA, QLoRA, and PEFT: curating datasets, choosing what to tune versus prompt, running the tune, and proving it beat the baseline, including when a smaller tuned model or distillation wins on cost and latency. Change behaviour and style, not facts.
Advanced retrieval
Hybrid search, reranking, query rewriting, and GraphRAG for connected knowledge; diagnosing retrieval failures that masquerade as hallucination. Most hallucination is a retrieval failure.
Agent design
Planning and tool-use loops, single versus multi-agent patterns, memory and state, cost and stopping conditions, and keeping autonomy inside guardrails. Autonomy inside guardrails, not without them.
Evaluation depth
Rigorous eval sets, LLM-as-judge with calibration, offline and online testing, and catching regressions before they ship. You can't fix what you can't measure.
Diagnose, go deep, measure, embed.
Diagnose
Bring a real quality problem; we work the framework to find the right approach, not the fashionable one.
Go deep
Hands-on modules on the chosen technique, built against your data and constraints.
Measure
Prove the advanced approach beat the simpler baseline on quality, cost, and latency; discard it if it didn't.
Embed
Practitioners leave with a repeatable method, not just one solved case.
Where it fits, the program runs learn-while-building on a real problem from your roadmap, so the output is production-grade work your team owns. To go deep alongside our engineers, see Dedicated Teams.
Choose techniques using evidence from your workload.
Technical judgement. Compare retrieval, fine-tuning and agent approaches against a representative evaluation set.
Quality trade-offs. Explain the quality, latency and running-cost differences behind the selected approach.
Operational readiness. Demonstrate failure handling and monitoring in the final practical exercise.
The deepest track of one upskilling path.
AI Upskilling and Training
The full path this advanced track sits at the end of.
Building LLM Applications
Ship a working LLM application first; this track goes deeper from there.
AI and GenAI Training
Apply generative AI to your role before going into practitioner depth.
AI for Beginners
The plain-language starting point for people new to generative AI.
Enterprise AI Literacy Program
Organization-wide, responsible understanding across every tier.
AI Strategy and Consulting
Focus that practitioner depth on problems worth solving.
What practitioners ask us first.
Engineers and ML practitioners who have already built and shipped an LLM application and need fine-tuning, advanced retrieval, agents, and rigorous evaluation. Newer engineers should start with Building LLM Applications.
No, and that is a core lesson. Much of the value is learning when not to fine-tune, because retrieval or prompting solves the problem faster and cheaper.
Yes. Learn-while-building on a real problem from your roadmap is our preferred mode, so the training produces work your team keeps.
Build a practitioner cohort around a real problem and the right approach to solve it.
Diagnose first, go deep on the technique that fits, and prove it beat the simpler baseline.