At a glance

Client: the software R&D arm of a global electronics conglomerate (name withheld under confidentiality).  Location: on-site, Bengaluru, India.  Format: 4-day classroom program, concept-first with live code demonstrations and supervised hands-on — a single trainer with a 30-engineer cohort.  Audience: experienced FE/BE engineers (5–10 years) new to AI, some with prior exposure.

What Was Covered

Sequenced so each day builds on the last — attention first, optimisation last

Day 1 · Transformers & Self-Attention

Why RNN/LSTM fall short and why transformers win; self-attention intuition (the "bank" and "Queen" disambiguation examples); embeddings and positional encoding; a worked Q/K/V pass; multi-head attention, feed-forward, residual + LayerNorm, and the stacked encoder — taught with worked classroom examples.

Day 2 · RAG, Hallucination & Guardrails

Why models hallucinate and how RAG grounds them; embeddings and vector search; chunking strategies (fixed vs recursive vs semantic) and their retrieval trade-offs. Live mutations of top-k and chunk size; breaking retrieval to prove the grounding link; and adding PII / decision-safety guardrails live, before/after.

Day 3 · Dialogue Management

Why an LLM alone isn't enough; the full ASR → NLU → Dialogue Manager → NLG → TTS pipeline; core patterns on a smart-TV assistant — slot filling, context handling, ambiguity resolution, corrections, and interruptions; why LangGraph (state machine vs prompt chaining); and a live LangGraph slot-filler that trainees extend.

Day 4 · Optimisation & Consolidation

The 50,000-user scaling scenario; KV-cache and quantisation (FP32 → INT8 → INT4) live; pruning, batching, flash attention, and speculative decoding — why and their trade-offs; the end-to-end optimisation story (1500 ms → 220 ms); and a full-pipeline recap tying RAG, guardrails, dialogue, and optimisation together.

How It Was Taught

Concept-first, with live code demonstrations and doubt-clearing built into each day. The hands-on model used live parameter mutation at the points that matter most — chunk size and top-k in RAG, adding a slot in LangGraph, toggling the KV-cache, comparing quantisation — so 30 engineers could watch behaviour change in real time under a single trainer, guided and supervised rather than free-form.

Outcome

A senior engineering cohort moved from AI-curious to able to reason about, build, and ground LLM/SLM features responsibly — understanding attention deeply enough to debug it, grounding answers with RAG and guardrails, shaping multi-turn conversations with LangGraph, and optimising for latency, memory, and cost at scale.

Related

← All Past Trainings

The full anonymized track record of delivered corporate AI trainings.

RAG Training →

Production retrieval: chunking, embeddings, pgvector, grounded answers, evals.

LangGraph Training →

Stateful, graph-based agents and dialogue management in depth.

Agentic AI Training →

End-to-end agent stack with LangGraph, tools, and evals.

Want this for your engineers?

From transformers to production LLM/SLM optimisation, scoped to your team's level and stack. Book a free intro call with ArjunThakur.dev.