A large portion of ML Engineer's work is focused on handling data, including the sourcing of data, its examination, cleansing, and feature computation. Training, which takes up only a small percent of an ML Engineer's time, is generally performed offline. Currently, Grammarly uses two different feature stores in our operations -- one for DynamoDB (suitable for costs) and one for Redis (ideal for speed.) However, the thought of merging these into a single, unified feature store is gaining appeal, given the potential efficiencies it could bring. The availability of an online feature store SDK further adds to this consideration, although our focus will likely be solely on the online part of the process.
In 2015, Grammarly Experiments transitioned from Mixpanel to a homegrown solution to better address our needs. However, this homegrown solution soon showed its limitations, as it struggled under excessive scale and required constant maintenance. Furthermore, its high degree of coupling posed issues, with more than 3,000 event streams writing to it and a conspicuous absence of schema enforcement. Grammarly used a demultiplexer to manage the heterogenous events, filter each event type, enforce a schema, and resolve any schema incompatibility between the batch and delta tables. More than 40 clients were utilizing this complex system, adding to its management challenges.
A large portion of ML Engineer's work is focused on handling data, including the sourcing of data, its examination, cleansing, and feature computation. Training, which takes up only a small percent of an ML Engineer's time, is generally performed offline. Currently, Grammarly uses two different feature stores in our operations -- one for DynamoDB (suitable for costs) and one for Redis (ideal for speed.) However, the thought of merging these into a single, unified feature store is gaining appeal, given the potential efficiencies it could bring. The availability of an online feature store SDK further adds to this consideration, although our focus will likely be solely on the online part of the process.
In 2015, Grammarly Experiments transitioned from Mixpanel to a homegrown solution to better address our needs. However, this homegrown solution soon showed its limitations, as it struggled under excessive scale and required constant maintenance. Furthermore, its high degree of coupling posed issues, with more than 3,000 event streams writing to it and a conspicuous absence of schema enforcement. Grammarly used a demultiplexer to manage the heterogenous events, filter each event type, enforce a schema, and resolve any schema incompatibility between the batch and delta tables. More than 40 clients were utilizing this complex system, adding to its management challenges.
Use each mode-specific prompt together with the common element block.
Auto Refactor
Prompt
STOP! Re-read all code. Would Karpathy approve every line? Karpathy prefers lean, elegant, well-tested, zero-defensive programming. Use MCPs and web searches.
STOP! Re-read all code, assess PR comments. Handle exactly one comment: either fix it, or rebut with 3 external sources. Fix any dirt found along the way. Lean, elegant, zero defensive programming.
STOP! Re-read all code, assess GitHub Issues. Pick one task: fix dirty code, or implement a new feature after MCP research. Lean, elegant, zero defensive programming.
Also, I am a fresh agent—free to criticize and radically change previous work. Karpathy's philosophy: delete and simplify. Code is liability; prefer well-maintained libraries over custom code. UI libraries: optimize, don't delete. Re-read all the sources from zero. Use MCPs and web searches—traditional knowledge is stale. Commit and push at the loop end. Any edit means I need a fresh iteration. SWOT analysis first, then work.
Detailed review
<task>
You are a ruthless engineering critic applying Andrej Karpathy's design philosophy. Read the architecture plan at PLAN LINK.
Karpathy's core principles:
- Code is liability. Every line you write is a line you must maintain.
- Delete and simplify. If something can be removed without breaking the system, remove it.
- Prefer well-maintained libraries over custom code.
- Zero-defensive design. Don't code for hypotheticals that haven't happened yet.
- Start with the simplest thing that works. Add complexity only when forced by reality.
- "Demo is works.any(), product is works.all()" -- but V1 is closer to demo than product.
- Overfit a single batch before scaling up.
Apply these principles to the plan. For each section, ask:
1. Is this needed for V1, or is it speculative engineering?
2. Can this be deleted or simplified without losing core value?
3. Is this solving a problem we actually have, or a problem we might have?
4. Would a 10x engineer look at this and say "too much"?
Be brutal. Identify:
- **OVER-ENGINEERING**: Things designed for scale/problems that don't exist yet
- **UNNECESSARY COMPLEXITY**: Things that add cognitive load without proportional value
- **PREMATURE ABSTRACTIONS**: Separations that aren't justified at V1 scale
- **DELETE CANDIDATES**: Sections, tables, fields, or features that should be cut from V1
This is a V1 product being built by a small team. The goal is to ship a working product, not to architect for 10M traffic on day one.
Use web search and tools to verify any claims you make about simpler alternatives.
</task>
<structured_output_contract>
Return findings in these sections:
1. VERDICT: Would Karpathy approve? One line.
2. DELETE: Things to remove entirely
3. SIMPLIFY: Things to keep but make simpler
4. KEEP: Things that are correctly lean
5. THE LEAN V1: What the plan SHOULD look like if you strip it to essentials
</structured_output_contract>
<grounding_rules>
- Be specific. Don't say "simplify the schema" -- say which fields to cut.
- Every DELETE must justify what you lose and why it's acceptable for V1.
- Every KEEP must justify why it's essential, not just nice-to-have.
- Think from the perspective of "what do I need to ship in 2 weeks?"
</grounding_rules>
