AI Learning System · Journey
Follow the learning flow from foundations to agent systems, assurance, and deployment decisions.
aiaLiving map · never “done”aiaLiving map · never “done”opens in a new tabgpuFoundationgpuFoundationopens in a new tabllmMain projectllmMain projectopens in a new tabuslFrom running to changing modelsuslFrom running to changing modelsopens in a new tabhnsReliable agent-system layerhnsReliable agent-system layeropens in a new tabctxWhat the model actually seesctxWhat the model actually seesopens in a new tabsecTrust and assurance layersecTrust and assurance layeropens in a new tabevlEvidence over assumptionevlEvidence over assumptionopens in a new tablclLocal deployment decisionlclLocal deployment decisionopens in a new tabcldCloud deployment decisioncldCloud deployment decisionopens in a new tabOrient
aia“What exists?”
Start with the map of the territory. Compare AI products and developer ecosystems with evidence before investing time in any track. The atlas is a living index, not a project to finish — every run through the loop updates it.
ecosystem models training inference runtime hardware cloud
aia.aserdargun.com opens in a new tabBuild · three parallel tracks
gpu“How does compute work?”
GPU Kernel Engineering — Kernel Atlas
Learn the machine the models run on: CUDA, Triton, GPU memory, profiling, and kernel optimization across a bilingual 12-week atlas.
CPU vs GPU GPU architecture VRAM memory bandwidth CUDA / Tensor cores FP32 / FP16 / BF16 / FP8 / INT8 / INT4 matrix multiplication CUDA kernels FlashAttention KV cache quantization multi-GPU tensor parallelism
gpu.aserdargun.com opens in a new tabllm“How do models run?”
LLM Runtime & Serving Atlas
Learn to ship models: runtime and serving solutions classified across seven architectural layers. Hardware and training both feed this track.
model architecture precision memory calculator runtime inference engine serving API benchmark
Ollama llama.cpp vLLM SGLang TensorRT-LLM Transformers MLX
llm.aserdargun.com opens in a new tabusl“How do models learn/change?”
Unsloth Studio Learning Atlas
Learn to change models, not just run them: Unsloth, LoRA, QLoRA, dataset engineering, and evaluation in a bilingual learning atlas.
pretrained model dataset tokenization LoRA QLoRA SFT DPO GRPO evaluation merged model LLM runtime
usl.aserdargun.com opens in a new tabgpu → llm hardware decides what you can serve usl → llm the models you train become what you serve llm → gpu inference pressure exposes the hardware limits
Engineer · agent system
hns“How do I turn model capability into a reliable agent system?”
Harness Engineering Observatory
Turn a served model into a dependable agent system. Compare harnesses, frameworks, runtimes, orchestration, sandboxes, verification, and observability with explicit evidence, freshness, and tradeoffs before choosing the surrounding system.
model capability context tools orchestration sandbox memory verification observability reliable agent system
hns.aserdargun.com opens in a new tabEngineer · context and assurance
ctx“What does the model actually see?”
Context & Knowledge Engineering
Design the information system that runs before model generation. Trace ingestion, chunking, embeddings, retrieval, citations, caching, memory, and tool use so the model sees grounded, governable context.
ingestion chunking embeddings vector DB retrieval citations cache memory tool use token budget
ctx.aserdargun.com opens in a new tabsec“Why should I trust this agent system?”
AI Systems Security Observatory
Trace delegated intent from model capability to auditable action. Examine identity, credentials, authorization, tool and sandbox boundaries, protected data, audit evidence, and incident recovery without collapsing assurance into a score.
model agent identity credential authorization tool sandbox data action audit incident
sec.aserdargun.com opens in a new tabevl“How do I know it works?”
AI Evaluation & Reliability Lab
Close the loop with evidence. Define evaluation contracts across output, trace, outcome, robustness, safety, and operations so every release becomes a deterministic, measured decision rather than an assumption.
golden set output trace outcome robustness safety operations release decision
evl.aserdargun.com opens in a new tabctx → sec context decides what is allowed in, security decides what is allowed out sec → evl assurance surfaces the questions evaluation must answer evl → ctx measurement feeds the next round of context design
Deploy · local and cloud in parallel
lcl“Which local lab should I buy?”
Local Compute Lab
Match open-weight model requirements to NVIDIA, AMD, and Apple devices. Compare workload, memory, privacy, power, noise, and existing equipment before choosing a local lab.
open-weight model workload memory NVIDIA / AMD / Apple privacy / power / noise local lab
lcl.aserdargun.com opens in a new tabcld“How do I operate this at scale?”
Cloud Provider Cost Comparison
Turn the stack into decisions: compare pre-tax USD costs of cloud services available from Türkiye and choose where the whole system runs. Cloud comes last on purpose — once you know what you deploy, infrastructure stops being abstract.
model vLLM Docker GPU instance cloud GPU load balancer autoscaling API
cld.aserdargun.com opens in a new tabReturn
The loop closes when what you learned returns to the system — the map you started with becomes the record of what you can now build, and the next pass through the loop starts deeper.