Week 19 — Fine-tuning trade-offs, GraphRAG¶

At a glance

Phase 5: Scale & Depth · Starts: Mon 01 Feb 2027 · Planned: 16h 45m

Build: Capstone: LoRA/QLoRA fine-tune experiment (Unsloth/TRL) vs prompt+DSPy baseline; GraphRAG spike

By day¶

System Design · 45 min · Active-active + conflict resolution (LWW, CRDTs) and DR drills → Multi-region, DR & failover w19-sd-1
DSA · 25 min · Longest Common Subsequence revisit (space-optimised) → Dynamic programming 2-D · resource w19-dsa-1
Python · 20 min · Rep: subinterpreters / concurrent.interpreters (3.14) experiment → Concurrency models: threads, processes, free-threaded 3.14t w19-py-1
Communication · 30 min · Grammar: Rhetorical devices: contrast, triads, rhetorical questions → Week 19 drills · day 1 w19-comm-1

Agentic AI · 60 min · Fine-tuning: when not to (prompting, RAG, DSPy first); LoRA/QLoRA mechanics; DPO vs GRPO — HF LLM course chapters → Fine-tuning: LoRA/QLoRA, DPO, GRPO — and when not to · resource w19-ai-1
DSA · 30 min · Maximum Subarray (Kadane) → Greedy · resource w19-dsa-2
Communication · 30 min · Vocabulary: Rhetoric and persuasion vocabulary → Week 19 drills · day 2 w19-comm-2

DSA · 25 min · Jump Game + Jump Game II → Greedy · resource w19-dsa-3
Architecture · 45 min · AI-native architecture II: agent reliability patterns — deterministic control flow, bounded autonomy, compensation → AI-native architecture: LLMs as system components w19-arch-1
Python · 20 min · Rep: refcounting, gc generations, memory layout of objects → CPython internals: bytecode, GIL, memory · resource w19-py-2
Communication · 30 min · Speaking drill: Persuasive talks → Week 19 drills · day 3 w19-comm-3

Agentic AI · 60 min · Advanced RAG: agentic RAG (not always better), GraphRAG, LazyGraphRAG — decide if capstone corpus benefits → Advanced RAG: agentic RAG, GraphRAG, LazyGraphRAG · resource w19-ai-2
Java/Spring AI · 30 min · JVM: Leyden AOT cache + compact object headers — measure startup/memory of the MCP server → JVM performance, GC & Leyden/AOT · resource w19-java-1
Communication · 30 min · Idioms & phrasal verbs: Persuasion idioms: make the case, win over → Week 19 drills · day 4 w19-comm-4

System Design · 45 min · Classic ML system design: recommendation & ranking funnel (candidate gen, ranking, re-ranking, features) → Classic ML system design: recommendation & ranking · resource w19-sd-2
DSA · 25 min · Gas Station → Greedy · resource w19-dsa-4
Python · 20 min · Rep: vectorise a hot loop with Polars instead of Python → Performance & profiling w19-py-3
Communication · 30 min · Writing: Blog post draft → Week 19 drills · day 5 w19-comm-5

Agentic AI · 120 min · Capstone build: QLoRA fine-tune a small model on 500 labelled triage examples with Unsloth/TRL; evaluate vs DSPy baseline; write ablation → Fine-tuning: LoRA/QLoRA, DPO, GRPO — and when not to · resource w19-ai-3
DSA · 30 min · Hand of Straights + Merge Triplets to Form Target Triplet → Greedy · resource w19-dsa-5
Java/Spring AI · 30 min · Virtual thread pinning diagnosis with JFR → Virtual threads, structured concurrency & scoped values w19-java-2
Communication · 30 min · Record & shadow: Persuasive talks → Week 19 drills · day 6 w19-comm-6

System Design · 45 min · Written design: coding agent / code-review bot — repo context, sandboxing, tool permissions, evals, cost → Design a coding agent / code-review bot · resource w19-sd-3
Architecture · 30 min · Write ADR 019 for capstone: prompt + DSPy over fine-tuning for v1 (with evidence from ablation) → AI-native architecture: LLMs as system components · resource w19-arch-2
Staff+ · 30 min · Decision journal: document 2 ambiguous calls from the capstone (options, info missing, reversibility, chosen path) → Decision-making under ambiguity w19-staff-1
Communication · 30 min · Soft skills + weekly review: Sponsoring and advocating for others → Week 19 drills · day 7 w19-comm-7
Review · 60 min · Explain-it-back: when to fine-tune vs RAG vs DSPy (3-min answer); flashcards on greedy proofs → Fine-tuning: LoRA/QLoRA, DPO, GRPO — and when not to w19-rev-1