LLM Engineering
Fine-Tune, Optimize, and Serve Large Language Models in Production
Fine-tune, quantize, and serve LLMs in production — on your own infrastructure.
AI Engineering treats the model as a black box behind an API. This course opens the box. When prompting and RAG aren't enough — when you need a model that speaks your domain, runs on your own hardware at a fraction of the cost, or serves thousands of requests within a latency budget — you are doing LLM engineering.
This is the most infrastructure-heavy course in the track. You'll learn when fine-tuning genuinely beats prompting and RAG (and when it doesn't), how to fine-tune efficiently with LoRA/QLoRA, how to compress models with quantization, and how to serve them in production with real throughput and latency. Self-hosted open-source inference is a first-class topic.
AI Engineering treats the model as a black box behind an API. That works — until it doesn't. When you need a model that speaks your domain, runs on your own hardware at a fraction of the cost, or serves thousands of requests within a latency budget, you are doing LLM engineering.
Most engineers reach for fine-tuning by reflex. Fewer know when prompting or RAG would serve better — and fewer still can actually fine-tune, quantize, and serve a model without a cluster and a cloud budget.
This is the most infrastructure-heavy course in the track. Self-hosted open-source inference with zero licensing cost is a first-class topic, not a footnote. The course is honest about cost and hardware throughout — the trade-offs are the curriculum.
Before
"I reach for fine-tuning when prompting doesn't work. I'm not sure if I need it, and I have no idea how to serve the result."
After
"I know when prompting/RAG wins and fine-tuning is wasted effort. I fine-tune with LoRA on modest hardware, quantize to fit my budget, and serve with real throughput targets."
- 1Knows when fine-tuning genuinely beats prompting and RAG — and when it's wasted effort
- 2Fine-tunes a useful model with LoRA/QLoRA on modest hardware, and quantizes it to fit a real budget with measured quality loss
- 3Self-hosts open-source inference (Llama, Mistral) at zero licensing cost — a first-class skill, not a footnote
- 4Serves models in production against a real throughput and latency target, not one request at a time
15 sessions · 30h across 4pillars · every session ends with a working, reviewable artifact. Expand any pillar for the session lineup.
LLM LiteracyHow models work, the lifecycle, decision framework · Week 1–2▼
- Session 1 — How LLMs Work, for Engineers
- Session 2 — The LLM Lifecycle: Pretraining, Fine-Tuning, Alignment
- Session 3 — Decision Framework: Prompt vs RAG vs Fine-Tune
Fine-TuningData, LoRA, infrastructure, evaluation · Week 3–5▼
- Session 4 — Fine-Tuning Data: The Real Work
- Session 5 — Parameter-Efficient Fine-Tuning (LoRA/QLoRA)
- Session 6 — Training Infrastructure & Frameworks
- Session 7 — Evaluating Fine-Tuned Models
Inference & ServingQuantization, optimization, production serving · Week 6–9▼
- Session 8 — Quantization & Model Compression (GPTQ/AWQ/GGUF)
- Session 9 — Inference Optimization (KV Cache, Batching, Speculative Decoding)
- Session 10 — Serving LLMs in Production I: Engines (vLLM/TGI)
- Session 11 — Serving LLMs in Production II: Scale & GPU Autoscaling
Production & CapstoneSelf-hosting, LLMOps, safety · Week 10–15▼
- Session 12 — Self-Hosting Open-Source Models (Llama, Mistral)
- Session 13 — LLMOps: Registry, Monitoring, Rollout
- Session 14 — Safety, Guardrails & Red-Teaming at the Model Layer
- Session 15 — Capstone: Fine-Tune, Quantize & Serve
Want the full session-by-session syllabus with objectives and deliverables? Request it here— we'll send it over.
| Delivery | Cohort-based, live, instructor-led |
| Class length | 2 hours per class |
| Cadence | 2 classes/week (4 hours/week) |
| Total length | 15 sessions · 30h over 7.5 weeks |
| Location | Dhaka + Global (remote seats available) |
| Prerequisites | AI Engineering (Course 4) or equivalent. Python, basic familiarity with model APIs. |
৳4,500
$69 global
Coming soonThis course launches soon. Join the waitlist for early access and launch-day pricing.
Do I need a GPU?▼
You'll need access to GPU compute for the fine-tuning modules. We provide guidance on cost-effective options (Colab, Lambda, RunPod, or your own hardware).
How much math is involved?▼
Comfort with math-adjacent concepts is expected, but we don't derive Transformer math. The focus is engineering — making models work in production.
Can I take this without AI Engineering (Course 4)?▼
Course 4 or equivalent experience is strongly recommended. You need basic familiarity with model APIs and prompt engineering.
Part of a longer track?
This course is part of the TIER B · ENGINEERING. Explore the full catalog to find related courses that build on each other.
View all courses →