Technical Glossary

An architecture style where applications are built as small, independent services. Often adds complexity without proportional benefits for most teams.

A single, unified codebase where all components run together. Often the right choice for teams under 50 engineers.

A cloud model where the provider manages servers. Despite the name, servers still exist—you just pay per-invocation instead of per-hour.

The cumulative performance and complexity cost of abstractions. Every layer between your code and the hardware has a price.

Shortcuts in code that create future maintenance burden. Unlike financial debt, it compounds exponentially and has no fixed interest rate.

Large Language Model. A neural network trained to predict text. Generates plausible-sounding output without understanding truth or intent.

When an AI confidently generates false information. Not a bug—it's how these models work. They optimize for plausibility, not truth.

Retrieval-Augmented Generation. Feeding an LLM relevant documents before asking a question. Reduces but doesn't eliminate hallucinations.

Training a model on your specific data to improve performance on your domain. More expensive than RAG but can improve accuracy.

The practice of crafting inputs to get better outputs from LLMs. More art than science—and models change frequently.

The maximum amount of text an LLM can process at once. Larger isn't always better—cost and quality often degrade at the edges.

AI systems that can take actions autonomously, not just generate text. Multiplies both capability and risk.

One billionth of a second. Light travels about 1 foot in a nanosecond—the fundamental limit Grace Hopper demonstrated with wire.

Time delay between request and response. Often more important than throughput. Physics sets hard limits that no amount of optimization can overcome.

Fast storage that keeps frequently-accessed data close to the CPU. L1 cache is ~1ns, RAM is ~100ns, SSD is ~100,000ns. The gaps matter.

The delay when a serverless function spins up from scratch. Can add 100ms-2s of latency, which kills user experience.

Data leaving a cloud provider's network. Often the hidden cost that makes cloud bills explode—$0.09-0.12/GB adds up fast.

Object-Relational Mapper. Translates between code objects and database tables. Convenient but often generates inefficient queries.

A performance antipattern where code issues one query, then N more for related data. Common with ORMs. Solution: joins or batch loading.

A self-contained, serverless database in a single file. Handles more load than most people think—often the right choice.

A powerful, reliable relational database with excellent features and a long track record. When in doubt, use Postgres.

A set of software development practices emphasizing iteration and collaboration. The principles are sound; the implementations often aren't.

Imitating the surface appearance of something without understanding the underlying principles. Common in tech methodology adoption.

A short daily meeting to synchronize a team. Often degenerates into status reporting theater that could be an async message.

Minimum Viable Product. The smallest thing you can ship to learn from real users. Often used as an excuse for shipping garbage.

The percentage of code executed by tests. High coverage doesn't mean good tests—you can have 100% coverage and still ship bugs.

The practice of having others examine your code before merging. Valuable for knowledge sharing; less effective for finding bugs than people think.

Automatic Speech Recognition. Converting audio to text. Accuracy claims are usually measured on clean TED talks, not your noisy real-world audio.

Word Error Rate. The standard metric for speech recognition accuracy. A 5% WER sounds great until you realize that's 1 wrong word per sentence.

Identifying who spoke when in an audio recording. Much harder than transcription—especially with overlapping speakers.

Voice Activity Detection. Determining when someone is speaking vs. silence. Sounds simple; edge cases make it hard.

Software as a Service. Subscription software hosted in the cloud. The pricing often hides significant per-seat costs.

How fast a startup spends money. When investors talk about "runway," they're calculating how long until your burn rate kills you.

The first major venture capital funding round, typically $2-15M. Comes with board seats and expectations of rapid growth.

Simple Agreement for Future Equity. A popular startup funding instrument that converts to equity later. Simpler than priced rounds but has traps.

The investigation process before an investment or acquisition. Technical due diligence often reveals surprises that kill deals.

Bulletin Board System. Pre-internet online communities accessed by dial-up modem. One user at a time. Where many of us learned to code.

A worldwide computer network that worked over phone lines before the internet. Store-and-forward messaging across the globe.

One of the first major online services (1969-2009). Email, forums, and chat before the web existed.