~ prathamesh

I build LLM agents that write, run, and fix their own infra automation.

Mostly I build the boring parts that keep them from paging you at 3am.

⌘K to jump around

about

I build LLM agent systems end-to-end — the parts that let a model write, run, and fix its own infra automation, and the guardrails that keep it from doing something dumb in production.

Right now I'm a product engineer at RunWhen, an early AI-SRE startup. I shipped our 0→1 agentic platform: an LLM that authors and self-corrects infra automation, owned from the React surface through the agent runtime to the backend.

Before that, ~3 years at Druva (via Josh Software) keeping enterprise backup at ~99.8% SLA and learning what actually breaks under real load.

Mostly I care about the boring engineering that stops the 3am page. Lately I'm poking at how far multi-agent systems scale before they fall over.

experience

4+ years across an early-stage startup and enterprise scale.

Product Engineer · RunWhen

Mar 2025 — present · Pune, IN (remote)

  • Built Build Mode 0→1: an agentic system where an LLM writes, runs, and self-corrects production infra automation. Picked a peer-agent architecture so non-engineers can ship Bash/Python tasks that run thousands of times a day.
  • Owned it full-stack — the React/Next.js tool-builder and chat surface, the agent runtime, and the backend behind it.
  • Grounded code-gen with RAG and built an LLM classifier that gates the agent's write access by task risk. Solo to beta in under a week.
  • Shipped in-VPC MCP server integration — the cheapest of four architectures I evaluated — making the platform cloud-agnostic for air-gapped, self-hosted sites.
  • Built a session-fork API for what-if branching of agent conversations — rewinds a deep-copied session to any invocation point without mutating the source.
  • Hardened the Kubernetes→PostgreSQL migration: root-caused a connection-pool exhaustion incident under load and closed a leak across 7 API endpoints. Cut workspace API latency >56%.

Software Engineer · Druva (via Josh Software)

Jan 2022 — Mar 2025 · Pune, IN

  • Owned platform services at ~99.8% SLA for high-throughput backup and metadata operations.
  • Led the monolith→microservices migration (REST + event-driven SQS/SNS) and drove Druva's first public API with its granular authorization framework.
  • Re-architected the backup path to cut resource use ~33% by killing redundant DB calls.
  • Root-caused a production incident cascading across seven services under live customer impact.
  • Wrote the HLD/LLDs that got multiple teams aligned on cross-cutting features.

open source

Contributions to developer tooling and infrastructure.

projects

Things I've built for fun and learning.

term

This resume, as an SSH TUI. Go + the Charm stack (wish, bubbletea, lipgloss). The web page you're reading is generated from the same data.

Kerr black-hole simulator

Real-time, physically-accurate renderer for a rotating (Kerr) black hole — per-pixel null-geodesic ray tracing in a GLSL shader. Three.js + Vite.

Time-Series Prediction

Compared Linear Regression vs Exponential Smoothing for forecasting — accuracy against compute trade-offs.

skills

Languages
PythonGoBash
Backend & data
FastAPIPostgreSQLRedisNeo4jClickHouseSQS/SNS
Infra & cloud
KubernetesDockerLinuxGCPAWSCI/CD
AI systems
LLM agentsmulti-agent systemsRAGMCPGoogle ADKLLM-as-judgecode-gen + review pipelines
Practices
system design (HLD/LLD)event-driven archmonolith→microservicesTDD

certifications

achievements

education

B.E. Information Technology

K.K. Wagh Institute of Engineering, Nashik · 2018 — 2022

CGPA 9.05/10

contact