# Shreyaan Seth

> Full-Stack Engineer and Engineering Lead focused on backend systems, AI infrastructure, and production reliability.

Canonical profile: https://www.shreyaan.me
Case studies: https://www.shreyaan.me/case-studies
Resume: https://www.shreyaan.me/Shreyaan-Seth-Resume.pdf
GitHub: https://github.com/Shreyaan
LinkedIn: https://www.linkedin.com/in/shreyaan-seth

## What Shreyaan builds

- React and Next.js product experiences, data-heavy interfaces, realtime UX, and accessible web applications.
- Production agent runtimes with tools, streaming, memory, delegation, approvals, retries, budgets, and evaluation.
- Durable event-driven workflows with persisted waits, human checkpoints, idempotent resume, and execution history.
- Multi-tenant retrieval, ingestion, memory, and enterprise access boundaries.
- Cloud and backend systems operated across virtual machines, containers, and serverless environments.

## Engineering profile

Shreyaan is a hands-on full-stack engineering lead. He personally builds core architecture and major product surfaces, then leads integration, production hardening, and scale. His deepest expertise is backend and AI infrastructure, supported by strong web fundamentals and systems knowledge.

## Case studies

Four long-form engineering case studies documenting real production decisions.

1. When an agent boundary stopped paying for itself
   https://www.shreyaan.me/case-studies/agent-runtime-architecture
   A production agent runtime serialized its own execution state into a second LLM purely to
   generate the final answer. Shreyaan wrote the RFC to remove that redundant synthesis
   boundary while preserving specialist-agent delegation, streaming, structured output, and
   step-level tracing. Removing the second mandatory LLM stage cut total response time from
   roughly 35s to 13s in some cases.
   Topics: agent runtimes, LLM orchestration, latency, observability, migration design.

2. Making long-running workflows survive process failure
   https://www.shreyaan.me/case-studies/durable-workflow-runtime
   A workflow could wait two days, but a deployment could destroy it, because continuations
   lived in an ephemeral application process. Shreyaan identified the failure model and pushed
   the runtime toward persisted waits, approval checkpoints, per-step execution state, and
   duplicate-safe resume claims. Core principle: workers should be disposable, workflow state
   should not be.
   Topics: durable execution, workflow engines, idempotency, distributed systems.

3. Moving tenant isolation from agent prompts into PostgreSQL RLS
   https://www.shreyaan.me/case-studies/postgres-rls-agent-isolation
   An AI SQL agent queried shared multi-tenant PostgreSQL tables. Shreyaan designed and
   implemented the PostgreSQL Row-Level Security layer end to end, including authenticated
   tenant context, database execution context, policies, role separation, and connection-pool
   safety. The model decides what query is useful; PostgreSQL decides which rows it may reach.
   Topics: PostgreSQL RLS, multi-tenancy, AI security, authorization.

4. Right-sizing an overprovisioned production PostgreSQL workload
   https://www.shreyaan.me/case-studies/production-postgres-rightsizing
   A production RDS instance billed about $988/month while averaging 0.8% CPU, ~3 connections,
   and using a tiny fraction of 30,000 provisioned IOPS. Shreyaan decomposed the bill by
   capacity dimension, found most waste in storage performance rather than compute, and resized
   conservatively rather than to the theoretical minimum. Observed RDS spend moved from about
   $987.50/month to $319.72/month (~68% lower).
   Topics: AWS RDS, capacity planning, cloud cost optimization, production risk.

## Verified operating evidence

- Operates 10 production services across virtual machines, containers, and serverless paths.
- Reduced a critical API path from approximately 7 seconds to 400 milliseconds.
- Leads architecture, delivery, reviews, and production debugging for a five-person engineering team.
- Removing a second mandatory LLM stage cut total response time from roughly 35 seconds to 13 seconds in some cases.
- Observed production RDS spend reduced from approximately $988/month to $320/month.

Company and client names are omitted. The technical decisions, ownership, and outcomes are kept intact.