The (Web) Reading Ruler
A browser userscript for saving and restoring a custom reading ruler across pages.
Engineer / storyteller / recovering hustler
Parin Katariya
I write about systems that scale, pauses that heal, and the strange little tools that make life on the internet feel more humane.
The pause has signal.
Career breaks, identity, spirituality, humor, small-town roots, and the work of becoming whole again.
Read Life EngineeringNotes for sharper builders.
Data structures, system design, interview patterns, and the mental models behind reliable software.
Read Engineering Agentic SystemsAgents need operating discipline.
Tool authority, harnesses, replay, approval, policy, and the distributed failure modes behind production agents.
Read Agentic Systems Projects & NotesUseful things, clearly explained.
Scripts, business concepts, product thinking, and practical artifacts from the workshop.
Read ProjectsFeatured
Best entry points
From Hustle to Heal: Why I Quit My High-Paying Job to Take a Break and Reclaim My Life
A candid, funny, deeply Indian essay about quitting a high-paying job to recover health and meaning.
Agentic Coding Changes the Bottleneck, Not the Responsibility
Why AI-assisted development moves senior responsibility toward framing, review, and integration.
Agentic Systems Are Distributed Systems With an Unreliable Planner
A 33-part engineering series on distributed systems, agentic systems, harness engineering, and production harness agents.
The Architecture Review That Actually Works
A review format that tests assumptions, failure modes, and operating ownership instead of style.
Evaluation Is the Product in AI Systems
Why durable AI products are built around eval loops, not model launch announcements.
Idempotency Is a Product Requirement
Why safe retries depend on product semantics, user-visible effects, and organizational ownership.
The Personal Operating System of a Principal Engineer
A practical operating model for attention, energy, writing, relationships, and technical leverage.
The Principal Engineer's Real Job Is Reducing Ambiguity
A field guide to turning vague technical anxiety into decisions, owners, and reversible next steps.
System Design
Scalability, load balancing, sharding, caching, and other system design fundamentals.
Archive
Everything on the desk
Life
Essays, pauses, and meaning
From Hustle to Heal: Why I Quit My High-Paying Job to Take a Break and Reclaim My Life
A candid, funny, deeply Indian essay about quitting a high-paying job to recover health and meaning.
Ambition Needs a Control Plane
How to keep ambition useful by adding constraints, feedback, and explicit tradeoffs.
Burnout Is a Throughput Bug
A systems model for burnout that treats recovery as capacity planning, not weakness.
The Career Value of Being Easy to Reason About
Why reliability, clarity, and low-surprise behavior compound into senior leadership trust.
The Dangerous Comfort of Being the Smartest Person in the Room
Why senior engineers must trade cleverness for systems that make other people sharper.
How to Build Taste as an Engineer
How engineers develop judgment by studying consequences, constraints, and user pain.
Parin's reflections -- A (fictional) story about myself
A fictionalized self-portrait about ambition, pause, spirituality, and finding a richer melody.
The Personal Operating System of a Principal Engineer
A practical operating model for attention, energy, writing, relationships, and technical leverage.
Om
A quiet opening prayer to mark the beginning of the blog.
What I Wish More Engineers Understood About Power
A clear-eyed model of power, trust, incentives, and responsibility in engineering organizations.
Why I Write Before I Decide
How writing turns vague opinions into inspectable engineering decisions.
Engineering
Systems, interviews, and code notes
Agentic Coding Changes the Bottleneck, Not the Responsibility
Why AI-assisted development moves senior responsibility toward framing, review, and integration.
Why AI Agents Need Smaller Interfaces Than Humans
How narrow tools, explicit contracts, and small write scopes make agents safer and more useful.
AI Code Review Needs a Threat Model
How to review machine-generated changes for subtle correctness, security, and maintainability risks.
Architecture Is Mostly Deciding What Must Not Change
A practical way to separate durable constraints from implementation details before design hardens.
The Architecture Review That Actually Works
A review format that tests assumptions, failure modes, and operating ownership instead of style.
Backward Compatibility Is an Organizational Skill
How compatibility survives when teams plan contracts, migrations, and deprecation paths together.
The Best API Design Happens Before the API Exists
A review method for designing APIs from lifecycle, failure, and ownership before endpoint shape.
The Case for Boring AI Infrastructure
Why queues, caches, eval stores, audit logs, and boring controls beat magic in production AI.
Why Data Quality Problems Look Like Engineering Problems
A taxonomy for separating data truth, pipeline behavior, model symptoms, and product expectations.
DDD
Notes from Parin's archive.
How to Disagree With a Staff Engineer Without Making It Political
A conflict model for keeping senior technical disagreement attached to evidence and risk.
Good Distributed Systems Are Designed Around Boredom
Why predictable recovery paths matter more than clever steady-state architecture.
Evaluation Is the Product in AI Systems
Why durable AI products are built around eval loops, not model launch announcements.
Feature Flags Are Distributed Systems in Disguise
The hidden consistency, cleanup, and ownership risks inside flag-driven delivery.
Feedback Loops Beat Model Launches
How feedback design determines whether AI systems improve, stagnate, or quietly drift.
The First 90 Days After a Technical Pivot
How to reset architecture, morale, and delivery after the product direction moves.
The Founder-Friendly Engineer Is Not a Yes-Man
How to support founder urgency while protecting the system from avoidable damage.
The Hidden Cost of Just One More Service
A principal-engineer checklist for deciding when a service boundary is worth its operational tax.
Hiring for Ownership When the System Is Still Moving
Signals for hiring engineers who can own ambiguous systems without needing stable ground.
Idempotency Is a Product Requirement
Why safe retries depend on product semantics, user-visible effects, and organizational ownership.
Incident Reviews Should Change Roadmaps, Not Just Runbooks
How to make incident reviews produce prioritization changes instead of better-formatted regret.
Influence Is Built Before the Meeting
Why senior influence comes from pre-wiring context, risks, and tradeoffs before rooms fill up.
Java DS Summary
A practical Java data-structures refresher for interview prep and daily problem solving.
The Difference Between Latency Budgets and Latency Wishes
A practical model for converting performance goals into owned budgets across a request path.
Latency, Tokens, and Trust in AI UX
Why AI user experience depends on cost, waiting, uncertainty, and visible recovery paths.
Leet code patterns summary
A compressed map of interview problem patterns: windows, pointers, graphs, dynamic programming, and more.
Leet code patterns with problems
LeetCode patterns paired with representative problems and descriptions.
How to Make Executives Trust Engineering Judgment
A communication model for translating engineering risk into business choices without theater.
Why Most Migrations Fail Before Code Starts
The migration risks that show up in ownership, sequencing, and rollback design long before implementation.
The ML Model Is Rarely the System Boundary
A systems view of model behavior, data pipelines, fallbacks, product semantics, and human review.
The New Skill Is Reviewing Machine Work
A review rubric for AI-generated code that focuses on intent, contracts, failure modes, and tests.
Observability Without Ownership Is Expensive Decoration
Why dashboards only become useful when somebody owns the question they are supposed to answer.
Observability Pillars
Notes from Parin's archive.
Why Platform Teams Accidentally Become Gatekeepers
How internal platforms drift from leverage engines into approval queues, and how to pull them back.
The Principal Engineer's Real Job Is Reducing Ambiguity
A field guide to turning vague technical anxiety into decisions, owners, and reversible next steps.
A Principal Engineer's Guide to Saying No
How to refuse unsafe work while preserving momentum, trust, and options.
Promotion Packets Are Written One Decision at a Time
How principal-level evidence accumulates through durable decisions, not last-minute brag documents.
Prompting Is Not the Hard Part of Agentic Development
Why context design, verification, decomposition, and ownership matter more than clever prompts.
The Queue Is Not the Architecture
How to reason about async workflows by ownership, ordering, failure semantics, and reconciliation.
The Reliability Work Nobody Wants to Fund
A framing for turning invisible reliability labor into risk reduction leaders can evaluate.
Startup Engineering Is the Art of Choosing Which Fires Stay Lit
A startup operating model for separating existential fires from noisy but survivable ones.
System Design
Scalability, load balancing, sharding, caching, and other system design fundamentals.
Technical Debt Is Often Unpriced Strategy Debt
A sharper way to discuss debt when the real issue is changing strategy, incentives, or ownership.
When Speed Becomes a Culture Smell
The difference between healthy urgency and speed used to hide poor choices.
Agentic Systems
Agents, tools, harnesses, and production control
Agentic Systems Are Distributed Systems With an Unreliable Planner
A 33-part engineering series on distributed systems, agentic systems, harness engineering, and production harness agents.
Adversary Agents
Adversary agents should create hostile fixtures and boundary tests, not become theatrical opponents in production workflows.
The Agent Action Lifecycle
A concrete lifecycle for turning uncertain agent intent into bounded, verified effects.
Agent Observability Should Follow Decisions, Not Tokens
A decision-centered observability model for debugging agent behavior after real side effects.
Agent State Should Live Outside the Model
A practical argument for storing task state, approvals, actions, and verification outside the conversation window.
Agents Need Read/Write Separation
Why useful agent autonomy starts by separating observation from mutation.
Ambiguous Outcomes Are the Core Distributed Systems Problem
The hard state in distributed systems is not failure, but not knowing whether the effect already happened.
Auditor Agents
Auditor agents review decision traces against policy and state, producing evidence-backed findings instead of vague confidence.
Backpressure Is a Product Boundary
Backpressure is where a system tells the truth about what it can accept, delay, prioritize, or refuse.
Blast-Radius Agents
Blast-radius agents decompose risky agent work into reviewable units without deciding the organization's appetite for risk.
Context Provenance Matters More Than Prompting
Why agent reliability depends on knowing where context came from, when it was captured, and what it is allowed to justify.
Control Plane vs Data Plane Failures
Separating control plane and data plane behavior keeps management failures from casually breaking committed work.
Deadlines Beat Timeouts
Timeouts protect individual calls; deadlines protect the whole promise the system made.
Failure Injection for Agents
Agent harnesses should inject ordinary tool and state failures so teams can verify behavior before production does it for them.
Golden Behaviors Beat Golden Text
Agent tests should preserve required behavior instead of freezing exact prose that can pass while actions regress.
Grade State, Not Speech
Agent quality should be judged by the state it leaves behind, because fluent explanations can hide unsafe actions.
A Harness Is a Controlled World
Agent harnesses earn trust by controlling the world around the model, not by asking the model to explain itself.
Harnesses Must Catch Overreach
Agent harnesses need explicit overreach checks because the most dangerous failures often look like extra helpfulness.
Human Approval Is a Protocol
How to make human approval precise enough to govern agent actions instead of decorating them.
Idempotency Is Durable Intent
Idempotency works when it preserves a decision across retries, crashes, and uncertain outcomes.
Leases, Ownership, and Fencing
A lease only says who probably owns work right now; fencing is what stops stale owners from writing after they lose it.
The Model Should Not Be the Policy Engine
Why agent systems should keep policy enforcement outside the model even when the model helps interpret intent.
Queues Are Not Reliability
A queue can preserve work and decouple timing, but it does not make processing correct by itself.
Reconciliation Is Where Correctness Becomes Real
A system is not correct because it accepted a command; it becomes correct when it can compare intent with reality and repair the gap.
Regression Suites Should Be Built From Bad Runs
The most valuable agent regression cases come from real failures converted into replayable behavior contracts.
Replay Agents
Replay agents reconstruct failed runs from durable traces so engineers can debug causes instead of arguing from summaries.
Replay Changes Everything
Replay turns production failures from arguments about logs into experiments against recorded inputs, decisions, and effects.
Replay Is the Difference Between Debugging and Guessing
Replay turns failed agent runs into inspectable infrastructure events instead of arguments about prompts.
Sagas Without Fantasy Rollbacks
A serious saga design treats compensation as forward recovery, not a pretend undo button.
Stop Conditions Are Safety Features
How explicit stop conditions keep agent systems from turning uncertainty into expanding side effects.
Test the Refusal Path
Safe agent systems need tested refusal paths, because declining action is a production behavior with side effects.
Tools Are Capabilities, Not Functions
A practical model for designing agent tools around authority, scope, and side effects.
Agents Are Unreliable Planners Attached to Real Tools
A production frame for treating agent plans as tentative while tool effects stay real.
Watcher Agents
Watcher agents are read-only harness components that notice unsafe drift and propose interventions without taking ownership of the work.
Projects & Notes
Tools, business, and practical artifacts
The (Web) Reading Ruler
A browser userscript for saving and restoring a custom reading ruler across pages.
The Agent Loop I Trust for Production Code
A practical loop for using coding agents without outsourcing engineering judgment.
The Cost Model Every AI Product Needs
A unit-economics lens for token cost, latency, quality, and human review in AI products.
The Quiet Power of Writing the Decision Record
How lightweight decision records create memory, alignment, and better reversibility.
Why Senior Engineers Should Read P&L Statements
How revenue, margin, cost, and risk make technical prioritization sharper.
Unit economics
Business metrics for engineers who want their technical decisions to connect to the P&L.