I am Srikanth Sastry. I am currently an engineering lead at Meta working on privacy infrastructure. Before Meta, I was a software engineer at Google in Cambridge, MA, and before that a postdoctoral associate with Nancy Lynch at CSAIL, MIT. Prior to that, I received my PhD from Texas A&M University.
My professional interests include distributed system design, software engineering processes, and building engineering teams. I also maintain a digital garden — a collection of evolving ideas and notes that complement my blog posts.
In a previous life, during my years in academia, I worked on distributed systems, networking, fault tolerance, and algorithm design and analysis. My list of publications are available on DBLP.
Recent Writing
All 65 posts →
The Guardrail Erosion Problem with AI Agents
AI coding agents erode the guardrails in your codebase. That erosion is structural, not accidental: it follows from the properties that make AI agents useful in the first place. You cannot eliminate it. But you can direct your finite human attention to the places where erosion is most dangerous, and let the right class of guardrail do the rest.
From the Garden
All 98 notes →AI Collapses the Economic Moat of Clean-Room Reimplementation
🌳The copyleft moat was never purely legal. It was economic: compliance was cheaper than reimplementation. AI collapsed that cost.
AI Security Cost Asymmetry Favors Closed Source
🌳AI equalizes the defender's security cost across open and closed source but varies the attacker's cost by source availability. This inverts Linus's Law.
Align Alerts to SEV Criteria
🌳Alerts should fire at or near the threshold where an SLO breach would occur, not well before.
Backward Compatibility for Leaky Abstractions
🌳When a framework leaks implementation details (like serializing arguments at schedule time but loading code from HEAD at execution time), changing a function signature breaks the assumption that old code calls old signatures.
Check If Concern Is Systemic
🌳Before resolving a concern, check whether it is isolated or systemic.
Command Control Misnomer
🌳"Command and control" is a misnomer for directive governance.
AI Reviewing AI: Shared Blind Spots
🌿AI models reviewing AI-generated code share systematic blind spots with the generator, creating gaps that neither side detects.
Confabulation Is Plausible
🌿AI agent confabulation is not random — it is plausible-looking wrongness constructed from pattern and proximity rather than knowledge.
Crisis Centralization Ratchet
🌿Tech companies centralize decision-making during crisis and almost never decentralize afterward.
Data Pipeline Is Achilles Heel
🌿The data pipeline is directive governance's Achilles heel, not the decision-maker's rationality.