Mandalas

756 days, sixteen identity shifts

A mandala is a 48-day cycle that ends with a capability you can prove. The 108-week course is 15 full mandalas plus a final 36-day graduation mandala - and every one of them changes who you are as an engineer.

15 mandalas × 48 days + 36-day graduation mandala = 756 days
The sixteen mandalas of the identity ladder
#MandalaWeeksIdentity gained
1 Build a reliable learning system Weeks 1–7 Disciplined learner
2 Learn the geometry beneath intelligence Weeks 7–14 Linear-algebra thinker
3 Numerical computation and differentiation Weeks 14–21 Numerical programmer
4 Probability, autodiff and statistical inference Weeks 21–28 Statistical reasoner
5 Experimentation meets retrieval Weeks 28–35 Retrieval and experimentation builder
6 Production reliability and measurable AI Weeks 35–42 Reliability-minded engineer
7 Distributed primitives and design reasoning Weeks 42–48 Distributed-systems implementer
8 Classic distributed systems Weeks 49–55 System designer
9 From system design to ML product science Weeks 55–62 ML systems practitioner
10 Build a mature experimentation platform Weeks 62–69 Experimentation engineer
11 Train, align and retrieve with LLMs Weeks 69–76 LLM training and retrieval engineer
12 RAG, agents and inference foundations Weeks 76–83 RAG and agent engineer
13 Become an inference systems engineer Weeks 83–90 Inference performance engineer
14 Advanced serving and reliable routing Weeks 90–96 Serving and routing engineer
15 Build the enterprise AI platform Weeks 97–103 Multi-tenant AI platform engineer
16 Graduation - portfolio, capstone and interview readiness Weeks 103–108 Portfolio-backed Staff AI Systems candidate
Full breakdown
Mandala 1 · Days 1–48 · Weeks 1–7

Build a reliable learning system

Disciplined learner

They replace vague plans to study AI with a working weekly process.

Key topics
  • Python engineering, typing, testing and async fundamentals
  • NumPy, vectorization and broadcasting
  • Git, notebooks, Makefiles and reproducible experiments
  • Docker, HTTP, APIs and Linux basics
  • Personal mathematics and systems diagnostics
  • Weekly journals, evidence tracking and retrospectives
  • Vector spaces, span, basis and dimension
Evidence to collect
  • course dashboard and repository
  • tested CLI project
  • numerical utility library
  • reproducible notebook
  • diagnostic gap map
  • containerized service
  • first mathematics problem set
Mandala 2 · Days 49–96 · Weeks 7–14

Learn the geometry beneath intelligence

Linear-algebra thinker

Embeddings, attention and matrix models become geometric objects. The learner can reason about them.

Key topics
  • Linear independence, rank and nullity
  • Inner products and dual norms
  • Orthogonality and projections, Gram–Schmidt
  • Coordinate transformations
  • Eigenvalues and diagonalization
  • PSD matrices and quadratic forms, Rayleigh quotients
  • SVD, pseudoinverse and low-rank approximation
Evidence to collect
  • projection and coordinate-transformation labs
  • PCA-style experiment
  • quadratic-form visualization
  • low-rank compression demo
Mandala 3 · Days 97–144 · Weeks 14–21

Numerical computation and differentiation

Numerical programmer

They learn to ask whether an algorithm is mathematically correct and numerically sane.

Key topics
  • Matrix norms and conditioning, perturbation analysis
  • LU, QR and Cholesky
  • Iterative solvers, power iteration and Krylov methods
  • Randomized SVD, tensor notation
  • Partial and directional derivatives
  • Gradients, Jacobians and Hessians, Taylor approximations
  • Chain rule on computational graphs
Evidence to collect
  • numerical decomposition library
  • solver benchmarks
  • perturbation experiment
  • randomized approximation lab
  • gradient and curvature visualizers
  • manual backpropagation worksheet
Mandala 4 · Days 145–192 · Weeks 21–28

Probability, autodiff and statistical inference

Statistical reasoner

They can explain why training and evaluation procedures work instead of treating them as imports.

Key topics
  • Computational graphs, forward/reverse-mode autodiff, JVPs and VJPs
  • Probability spaces, conditional probability and Bayes
  • Random variables and distributions
  • Covariance and correlation
  • LLN, CLT and delta method
  • MLE and MAP, estimator bias and variance
  • Hypothesis testing, Type I and Type II errors
Evidence to collect
  • tiny autodiff engine
  • probability simulations
  • covariance analysis
  • Monte Carlo convergence demo
  • estimator comparison
  • statistical-test implementation
Mandala 5 · Days 193–240 · Weeks 28–35

Experimentation meets retrieval

Retrieval and experimentation builder

They can now build the primitives behind search, experimentation and traffic control.

Key topics
  • p-values, confidence intervals, bootstrap and permutation tests
  • power, MDE and sample-size calculations, multiple-testing correction
  • cosine similarity, exact and memory-bounded KNN
  • streaming Top-K and dynamically changing scores
  • two-stage retrieval and reranking
  • sliding-window and token-bucket rate limiting
  • LRU and TTL caching
Evidence to collect
  • bootstrap CI package
  • experiment-size calculator
  • vector-search mini-lab
  • Top-K service
  • reranking prototype
  • rate-limiter module
  • cache simulator
Mandala 6 · Days 241–288 · Weeks 35–42

Production reliability and measurable AI

Reliability-minded engineer

They stop thinking only about model output and start thinking about trust, failure, observability and operational behavior.

Key topics
  • text and sentence-aware chunking
  • PII span merging, regex and NER-based redaction
  • Expected Calibration Error, Brier score and reliability diagrams
  • NDCG, MRR and Recall@K
  • thread-safe inference batching
  • retries, exponential backoff and jitter, circuit breakers
  • Count-Min Sketch and Bloom filters
Evidence to collect
  • chunking library
  • safe redaction utility
  • calibration dashboard
  • retrieval evaluator
  • concurrent batcher
  • fault-tolerance toolkit
  • probabilistic data structures
Mandala 7 · Days 289–336 · Weeks 42–48

Distributed primitives and design reasoning

Distributed-systems implementer

They can move from implementing one process to reasoning about many machines.

Key topics
  • HyperLogLog, consistent hashing, rendezvous hashing
  • weighted load balancing, distributed unique IDs
  • leaky-bucket and sliding-window counters
  • requirements and constraint gathering, APIs and data modelling
  • observability, rate limiting at distributed scale
Evidence to collect
  • cardinality estimator
  • routing lab
  • load-balancer module
  • platform-primitives library
  • system-design document template
  • first complete design exercises
Mandala 8 · Days 337–384 · Weeks 49–55

Classic distributed systems

System designer

They stop naming technologies and start defending architecture.

Key topics
  • scaling reads and writes, caching and replication, contention
  • large blob storage and CDNs
  • orchestration versus choreography, workflow-state ownership
  • backward-compatible schema evolution
  • URL shorteners, Dropbox, Ticketmaster, News Feed, WhatsApp, LeetCode-style judges
Evidence to collect
  • scaling-pattern notes
  • blob-storage design
  • workflow-state architecture
  • schema-migration plan
  • multiple end-to-end design documents
Mandala 9 · Days 385–432 · Weeks 55–62

From system design to ML product science

ML systems practitioner

A production ML system begins with a decision and a dataset, rather than model training alone.

Key topics
  • Uber and web crawlers, ad-click aggregation, payments
  • business-objective formulation, model-versus-non-model decisions
  • labeling and weak supervision, data augmentation
  • leakage and contamination, experiment tracking and versioning
  • baselines, slice-based offline evaluation, confidence intervals
  • shadow and canary deployment, bandits
Evidence to collect
  • ML project charter
  • data-quality checklist
  • leakage audit
  • evaluation harness
  • rollout plan
Mandala 10 · Days 433–480 · Weeks 62–69

Build a mature experimentation platform

Experimentation engineer

They can tell when a metric moved and when the evidence supports changing the product.

Key topics
  • A/B-test significance calculation, absolute and relative lift
  • pooled standard errors, confidence intervals
  • paired bootstrap model comparison
  • Sample-Ratio Mismatch detection, chi-square diagnostics
  • feature-flag engines, deterministic sticky assignment
  • monitoring and drift detection, production incident analysis
Evidence to collect
  • analyze_ab_test
  • bootstrap comparison package
  • SRM detector
  • feature-flag service
  • experiment-assignment service
  • monitoring specification
  • incident postmortem
Mandala 11 · Days 481–528 · Weeks 69–76

Train, align and retrieve with LLMs

LLM training and retrieval engineer

They understand the whole upstream and retrieval lifecycle rather than treating the LLM as a remote magic endpoint.

Key topics
  • transformer and tokenizer anatomy, training-data cleaning and deduplication
  • SFT, LoRA, QLoRA and DoRA
  • RLHF, DPO, ORPO and GRPO
  • DDP and FSDP, tensor and pipeline parallelism, ZeRO
  • HNSW, IVF-PQ and product quantization
  • hybrid retrieval, query rewriting, parent–child and multi-vector retrieval
Evidence to collect
  • model-anatomy notes
  • dataset card
  • fine-tuning plan
  • preference-evaluation plan
  • distributed-training architecture
  • vector-index benchmark
  • retrieval lab
Mandala 12 · Days 529–576 · Weeks 76–83

RAG, agents and inference foundations

RAG and agent engineer

They can build an agent and explain where its latency, errors, context and cost come from.

Key topics
  • BM25 + dense + reranker fusion, Reciprocal Rank Fusion
  • evaluation by query slice
  • prompt templates and version registries
  • conversation and long-term memory, context-window optimization
  • ReAct, plan-and-execute, tool calling, agent guardrails
  • GPU compute-, memory- and overhead-bound regimes, FlashAttention
  • roofline models and arithmetic intensity
Evidence to collect
  • retrieval-fusion service
  • regression report
  • prompt-management tool
  • memory service
  • agent prototype
  • GPU performance primer
  • arithmetic-intensity worksheet
Mandala 13 · Days 577–624 · Weeks 83–90

Become an inference systems engineer

Inference performance engineer

They can diagnose inference rather than simply complain that it is slow.

Key topics
  • PagedAttention, KV-cache block tables, vLLM deployment
  • scheduler and block-manager internals, running/waiting metrics
  • latency histograms, Prometheus and Grafana
  • request-rate sweeps and saturation analysis
  • SGLang, RadixAttention and prefix reuse
  • Orca-style continuous batching, chunked prefill
Evidence to collect
  • live vLLM deployment
  • internal architecture notes
  • metrics schema
  • serving dashboard
  • benchmark methodology
  • vLLM/SGLang comparison
  • scheduling report
Mandala 14 · Days 625–672 · Weeks 90–96

Advanced serving and reliable routing

Serving and routing engineer

They can optimize a model and a multi-provider inference service.

Key topics
  • continuous batching trade-offs, FP8, AWQ and GPTQ
  • KV-cache compression, speculative decoding, Medusa and EAGLE
  • StreamingLLM and KV eviction
  • DistServe, Splitwise and Mooncake, prefill/decode disaggregation
  • model routing, canary assignment and kill switches
  • PII-safe traces and logs, async fan-out, partial failures, token-stream multiplexing
Evidence to collect
  • quality-versus-throughput benchmark
  • latency and memory lab
  • disaggregated serving architecture
  • model-router service
  • safe tracing layer
  • resilient aggregator
Mandala 15 · Days 673–720 · Weeks 97–103

Build the enterprise AI platform

Multi-tenant AI platform engineer

This is where “tensor” finally becomes “tenant.”

Key topics
  • distributed token budgets, request prioritization
  • DAG workflow execution, idempotency
  • per-tenant concurrency limits, leader election
  • distributed work queues, dynamic configuration
  • model health monitoring, registries and promotion
  • semantic and embedding caches, prompt-injection detection
  • exact cost attribution, capstone architecture planning
Evidence to collect
  • quota and priority service
  • workflow engine
  • tenant limiter
  • HA lab
  • queue/configuration platform
  • model-operations pipeline
  • cost/safety platform
  • complete capstone design document
Mandala 16 · Graduation · Days 721–756 · Weeks 103–108

Graduation - portfolio, capstone and interview readiness

Portfolio-backed Staff AI Systems candidate

They no longer need to tell someone they “know AI systems.” They can provide the repository, traces, benchmarks, design documents and failure tests.

Key topics
  • complete the capstone core API
  • build the retrieval or agent flow
  • install evaluation, tracing and guardrails
  • run load and failure tests, tune latency and cost
  • produce architecture and benchmark reports
  • complete coding, system-design and behavioral mocks
  • publish the portfolio, perform a full retrospective
  • create a 30-day career plan
Evidence to collect
  • repository, traces, benchmarks, design documents and failure tests
  • published portfolio release
  • mock-interview scorecard
  • 30-day job-search plan