AI Product Teardowns
For engineers and researchers who build AI systems
How real AI products work under the hood. Architecture decisions, tech stacks, and key engineering insights.
GitHub Copilot
Tech Stack
How it works
Copilot is no longer just inline autocomplete. The current architecture spans IDE chat and completions, local agent mode, and a cloud agent that researches a repository, drafts a plan, edits a branch, runs tests in an ephemeral GitHub Actions environment, and can open or update pull requests.
Key Insight
The moat has shifted from completion latency to workflow ownership: GitHub controls issues, branches, reviews, Actions, permissions, usage metrics, and enterprise policy around the agent.
ChatGPT
Tech Stack
How it works
ChatGPT has become a front door for multiple model and tool paths: fast everyday chat, reasoning-heavy work, web search with citations, deep research reports, file/data analysis, image creation, scheduled tasks, connected apps, and Codex for software work.
Key Insight
The product advantage is orchestration. ChatGPT packages model choice, memory, files, browsing, app permissions, agent execution, and creative tools into one consumer workflow instead of exposing each capability as a separate developer primitive.
Midjourney
Tech Stack
How it works
Midjourney has expanded from Discord-first image generation into a web creative workflow with image editing, reference-driven generation, and video animation from generated or uploaded stills. The public product is still a closed creative model stack rather than an API-led platform.
Key Insight
Midjourney's differentiation remains taste and creator workflow. Its strongest users are not buying a generic diffusion endpoint; they are buying aesthetic defaults, iteration speed, references, and community-shaped discovery.
Cursor
Tech Stack
How it works
Cursor now presents itself as a coding agent across desktop, CLI, cloud, review, and autocomplete surfaces. The workflow ranges from fast Tab predictions to targeted edits, repository-aware agent plans, shell/tool use, and pull-request handoff.
Key Insight
Cursor's product wedge is the autonomy slider: developers can stay in low-friction autocomplete, ask for scoped edits, or let an agent plan and change larger parts of the codebase without leaving the editor workflow.
Perplexity
Tech Stack
How it works
Perplexity has evolved from cited answer search into a search-and-agent stack: consumer answers, developer APIs for Sonar/search/agents, and Comet, a Chromium-based browser with an assistant that can summarize, research, email, shop, and act across web pages.
Key Insight
The core design bet is that search becomes an action surface. Citations build trust for answers, while Comet tests whether the browser itself can become the execution environment for everyday web tasks.
HyperCLOVA X
Tech Stack
How it works
HyperCLOVA X is now a Korean-first model family rather than one monolithic LLM. NAVER positions THINK for heavier reasoning, SEED variants for efficient deployment, and omni/vision/audio capabilities for multimodal Korean workflows through CLOVA Studio and NAVER service integrations.
Key Insight
The defensible layer is local context: Korean data, token efficiency, cultural nuance, domestic customer cases, and distribution through NAVER Cloud and NAVER-owned products matter more than frontier-model size alone.
Claude Code
Tech Stack
How it works
A terminal-first coding agent: a model wrapped in a harness of tools (bash, file read/edit, search), project memory in a CLAUDE.md checked into the repo, skills as on-demand instruction folders, hooks that intercept tool calls, and subagents that fan work out. The same harness ships as the Agent SDK and runs headless in CI.
Key Insight
The product is the harness, not the model. The layer that compounds — CLAUDE.md, skills, hooks — lives in git alongside your code, so the agent gets better per-repository over time. A chat window can't accumulate that.
DeepSeek
Tech Stack
How it works
A cost-first frontier architecture built under GPU export constraints: a highly sparse mixture-of-experts that activates a small fraction of parameters per token, Multi-head Latent Attention that compresses the KV cache, and FP8 training. R1 then showed reasoning could be trained with large-scale RL and distilled down into small open models — all released under MIT.
Key Insight
DeepSeek treated cost as the product. The architecture exists to make frontier-class tokens radically cheaper, and the open weights plus aggressive API pricing reset the entire industry's economics — every closed lab now prices against it.
Solar LLM
Tech Stack
How it works
Korea's startup route to a frontier model: Solar began with depth up-scaling — stacking layers of a proven smaller model, then continuing pretraining — and grew into the Solar Pro line, the first Korean family Artificial Analysis recognized as frontier-class. Upstage sells it as a stack with Document AI (parsing, OCR), and holds one of the five MSIT sovereign-AI mandates.
Key Insight
Solar proves you don't need hyperscale compute to stay relevant — architectural pragmatism plus a wedge enterprises actually pay for. The chat model gets the benchmarks; the document stack gets the contracts.
Kling
Tech Stack
How it works
A diffusion-transformer video model shipped inside Kuaishou's short-video ecosystem rather than as a standalone demo. Generation is a feature in a creator funnel with hundreds of millions of viewers: successive model generations improved motion and physical consistency, monetized through creator subscriptions and an API.
Key Insight
The moat is distribution, not the diffusion transformer. Rivals ship impressive demos; Kling ships to an audience that immediately posts, monetizes, and feeds engagement data back — which is why it out-earns peers with comparable models.
Groq LPU
Tech Stack
How it works
Inference silicon that throws out the GPU playbook: no HBM and no dynamic scheduling. Model weights live in on-chip SRAM spread across many chips, and a compiler schedules every cycle statically — so latency is deterministic and single-request token speed is extreme. Sold as the GroqCloud API rather than as boards.
Key Insight
Groq bet that inference is a compiler problem, not a memory-bandwidth problem. Determinism turns hardware into a predictable pipeline — and the 3M+ developers on GroqCloud, not chip specs, are what cleared the catalog's hardware evidence bar.
Tesla FSD
Tech Stack
How it works
From v12, FSD replaced hundreds of thousands of lines of hand-written control heuristics with a single end-to-end network from camera pixels to steering and pedals. It trains on fleet video curated by a data engine, validates candidate models silently in shadow mode against human drivers, and ships over-the-air to millions of cars.
Key Insight
The fleet is the lab: every mile generates training data and every OTA push is a deployment. The contested bet is vision-only — trading sensor redundancy for a data flywheel that no lidar-equipped fleet can match in scale. This is the clearest industrial preview of the embodiment wave.
More teardowns coming soon
We're analyzing how every product in our database uses AI under the hood