The Advanced Project Portfolio — what to actually build

The heart of your request. The rule before the list: one flagship done to completion beats ten half-projects — and "completion" means the four-part ship: repo + evals/numbers + writeup + live demo. Here's the tier list, each project mapped to the course modules it proves:

Tier S — flagships. Pick ONE and go deep:

1. "My 8B beats [frontier model] on X" — the full specialist pipeline (Modules 2–3–10): build a dataset with your Topic 17 synthetic pipeline + Topic 18 curation, QLoRA + DPO it, then rigorously evaluate against a frontier model on a firewalled golden set — and publish model, dataset, eval harness, and the writeup with the numbers. Task choice is strategic: pick something underserved where beating the giant is achievable — Urdu/Roman-Urdu tasks (your unfair advantage: frontier models are genuinely weak there and you can judge quality natively), SQL generation for a schema family, domain-specific structured extraction. This project is the single strongest "LLM engineer" credential that exists below actual lab experience, because it proves the entire stack.

2. An inference-optimization project with published benchmarks (Module 4 spike): implement speculative decoding or a KV-cache/quantization technique from scratch (or contribute one to llama.cpp/vLLM), or build a "quantization quality atlas" — Q8→Q2 across 5 models × 4 task types, with a public results table. The systems-engineering audience for this is exactly the audience that hires at infra startups.

3. An agent scored on a public benchmark (Modules 7/10): a SWE-bench-lite or τ-bench submission with a serious trajectory-analysis writeup — where it failed, why, what fixed it. A middling score with brilliant failure analysis signals more engineering maturity than a demo that "usually works."

4. A rigorous eval suite for an underserved domain (Module 10 spike): e.g., a contamination-resistant Urdu benchmark, or a domain eval (regional legal/medical/financial tasks) with meta-evaluated LLM judges (Topic 75's full rigor). Evals are the most supply-constrained skill in the industry right now — labs and startups alike hire eval-obsessed people on sight, and almost nobody's portfolio demonstrates it.

5. A production-grade RAG system with ablations (Module 6): hybrid + reranker + golden set — and the part that makes it advanced: the ablation study (chunk-size sweep, reranker on/off, recall@k for each) published as a table. "I built RAG" is Tier B; "here's what each pipeline stage is worth, measured" is Tier S.

Tier A — strong supporting pieces (2–3 of these around your flagship): an MCP server for a real tool (early on a spreading standard = discoverable); a multi-tenant LoRA serving demo (one base, hot-swapped per-customer adapters — Topic 22's wardrobe, which very few people have actually built); your synthetic-data pipeline extracted as a reusable OSS tool (generation + the filtering stack); a local AI app shipped to real users (hardware-tiered, Topic 68's playbook); and the sleeper: a from-scratch reimplementation with a writeup — DPO or GRPO in 300 lines, Flash-Attention tiling in Triton, a tiny inference engine. "I built X from scratch and here's what I learned" is beloved by exactly the engineers who do hiring.

Tier B — commoditized; never flagships: generic chatbots, chat-with-your-PDF, framework-wrapper demos, prompt collections. Fine as learning exercises (you did several in this course); invisible as portfolio.

The meta-rule tying it together: choose projects that generate writing material and numbers — because Topic 88's content engine feeds on them, and Module 10's discipline is your watermark on every one.

---