System design, AI engineering, databases, networking, concurrency and more, written for engineers building real trading and fintech systems — part of the Algomi Suite.
Turning the mechanics of LLM inference into a capacity plan: TTFT and tokens-per-second targets, how many GPUs a model needs, tensor and pipeline parallelism, prefix caching, speculative decoding, serving engines, streaming, and autoscaling.
Created 2026-09-29 · Updated 2026-10-15 · AI Solutions Architecture & AI Engineering
How approximate nearest neighbour indexes (HNSW, IVF, product quantisation) work and trade recall for speed, then the production concerns the basic RAG guide skips: permission-aware retrieval, multi-tenancy, real-time updates, hybrid search, and RAG evaluation.
Created 2026-09-29 · Updated 2026-10-15 · AI Solutions Architecture & AI Engineering
The design patterns behind reliable LLM agents: ReAct and plan-and-execute loops, agents as explicit state machines, the memory taxonomy, multi-agent trade-offs, and the safety rules for tools that change the world.
Created 2026-09-29 · Updated 2026-10-15 · AI Solutions Architecture & AI Engineering
Defending LLM applications against prompt injection and token exhaustion, keeping them up with provider failover, circuit breakers and graceful degradation, and keeping the bill predictable with budgets, routing, and caching.
Created 2026-09-29 · Updated 2026-10-15 · AI Solutions Architecture & AI Engineering
An eight-step framework for AI system design interviews, applied to eleven case studies: enterprise ChatGPT, coding assistant, AI search, knowledge assistant, support AI, document search, AI gateway, multi-model platform, agent platform, and evaluation and observability platforms.
Created 2026-09-29 · Updated 2026-10-15 · AI Solutions Architecture & AI Engineering
Prompts tell an agent what to do; a harness decides what it knows, what it may touch, and what counts as done. Five subsystems, the session lifecycle, and a working Java shift-handover model, told as a hospital night ward.
Created 2026-10-02 · Updated 2026-10-15 · AI Solutions Architecture & AI Engineering
How to turn repeated prompting into a controlled loop: goal loops, a separate checker, retry budgets, state kept outside the loop, reusable skills, and a tool gateway with permissions, timeouts and idempotency. Told as a child doing homework.
Created 2026-10-02 · Updated 2026-10-15 · AI Solutions Architecture & AI Engineering
When one loop is not enough: model the workflow as a graph, fan work out and back in, isolate parallel agents, roll back on failure, gate risky steps on a human, and measure the harness with ablation tests. Told as building a house.
Created 2026-10-02 · Updated 2026-10-15 · AI Solutions Architecture & AI Engineering
Why a company with hundreds of MCP servers needs a gateway, and how a control plane, a data plane, tool-level authorization, progressive discovery and response projection work, using Uber's published MCP Gateway as a case study.
Created 2026-10-04 · Updated 2026-10-15 · AI Solutions Architecture & AI Engineering
How to stop one company's documents leaking into another company's answers: tenant isolation at ingestion, retrieval, context assembly and caching, and how to choose between a shared index, namespaces and dedicated indexes.
Created 2026-10-04 · Updated 2026-10-15 · AI Solutions Architecture & AI Engineering
The agent loop, designing tools the model can actually use correctly, reactive versus planned execution, memory and compaction, multi-agent orchestration, and the guardrails that keep an acting AI safe.
Created 2026-09-18 · Updated 2026-10-15 · AI Solutions Architecture & AI Engineering
Why a demo that works for one user can collapse at scale — prefill vs. decode, the KV cache, continuous batching, PagedAttention, quantization, caching, and a worked cost estimate.
Created 2026-09-18 · Updated 2026-10-15 · AI Solutions Architecture & AI Engineering
Why a fluent AI answer can still be wrong — chunking, embeddings, hybrid search and reranking, and when retrieval beats fine-tuning or a bigger context window.
Created 2026-09-18 · Updated 2026-10-15 · AI Solutions Architecture & AI Engineering
Why a fluent AI answer isn't proof of success — evals, traces, groundedness, graders, and the release checks that catch failures a good average score would hide.
Created 2026-09-16 · Updated 2026-10-15 · AI Solutions Architecture & AI Engineering
