Anuj Gupta

AI platform architect

I'm Anuj Gupta. I've spent 14 years in software engineering, most of it on the systems other engineers build on. Along the way I led a 1,200-server cloud migration, designed an AWS data lake for ML and analytics, and architected the developer platform 1,000+ developers use today. These days it's agentic AI running in production. Master's in Computer Science, Rice University (2026).

01 Now building

My product · in pilot

Seronera

Live at seronera.dev · design partners invited

Seronera runs next to each developer's AI coding agent and takes every change through six gated stages. The agent drafts the intent, spec, plan and code, and a named person approves each one before the change moves on. Every approval is a commit in the team's own git history, and code never goes to a Seronera server.

  1. 1Plan
  2. 2Design
  3. 3Build
  4. 4Test
  5. 5Deploy
  6. 6Maintain

My product · early access

AgentGuardian

Self-hosted · private early access

AgentGuardian is a self-hosted control layer for AI agents in production. Every governed tool call gets a decision before it runs: allow, deny, or hold for a person to approve. When an agent fails, it opens an incident and retries, quarantines or escalates to on-call without a redeploy. Each decision goes into a hash-chained audit log.

  1. 1Connect
  2. 2Enforce
  3. 3Recover

My product · private pilot

Triadic

For people who connect people

Connect people on what they value. Build triangles that hold. Triadic is an iPhone app that remembers what each person cares about and needs, then shows you which two people you know should meet, and why. Notes stay on your phone unless you choose to send one online.

02 Selected work

Metrics are derived from production implementations. Company and internal project names have been omitted for confidentiality.

Agentic AI · FinOps

Autonomous FinOps platform

  • ~$2.1M projected annual savings
  • 150+ AWS accounts

Cloud waste and misconfiguration pile up faster than a small team can review them. I architected a multi-agent platform on Amazon Bedrock that finds them, proposes fixes, applies what a person approves, and returns a week later to score its predictions against actual savings. The model never holds credentials, and nothing irreversible runs without human sign-off.

Agentic AI · Security

Agentic vulnerability triage and remediation

  • 80% of triage automated
  • 60% MTTR reduction

Security triage was consuming four full-time engineers. I architected the platform (Bedrock, AgentCore, LangGraph) that automated most of it, with the same rule as everything I ship: the model proposes, deterministic code validates, a person authorizes.

Public write-up in progress.

Platform · Kubernetes

Kubernetes platform, built in-house

  • ~$6M vendor contract replaced
  • 140+ clusters across 150+ accounts

We replaced a commercial cluster-management product with a CNCF open source stack and migrated 140+ production EKS clusters to it in four months, with zero rollbacks. Fleet-wide upgrades went from weeks of manual effort to a single Git commit.

03 How I think

  1. The hard part of multi-agent AI is the substrate, not the agents. Agents take weeks. Audit trails, cost attribution, and deterministic testing take months.
  2. Model output is data, never instructions. A parser acts only on schema-validated output, and the model never holds credentials.
  3. Governance works when it's built into the platform. Once evals and guardrails were the default rather than an add-on, teams adopted them without being asked.
  4. FinOps done right improves architecture. Consolidating to hit a cost number is what made the estate governable at all.

04 Contact

I'm open to conversations about enterprise AI platforms and Staff, Principal, or Distinguished architect roles.