Storage & performance

Storage benchmark agent

Describe a storage workload in plain English. The agent generates configs for fio, vdbench, iozone, and iometer, runs all four in parallel, parses IOPS / throughput / latency / CPU, and produces a full side-by-side comparison report saved to disk.

  • Natural language → fio + vdbench + iozone + iometer configs automatically
  • All 4 tools run concurrently via ThreadPoolExecutor — no waiting
  • Parses IOPS, throughput, P99 latency, and CPU from each tool
  • Full comparison report saved as JSON + text on every run
  • Dry-run mode — full simulation with no tools or API key needed
  • GitHub Actions CI included — green badge on every push
  • Rate-limit safe — 5s throttle keeps under Gemini free tier (15 RPM)
LangGraph Gemini 1.5 Flash fio vdbench iozone iometer ThreadPoolExecutor Python 3.9+
Agent flow
Natural language prompt
"70/30 R/W, 128K, 32 threads, 60s"
Planner (Gemini)
generates 4 configs
↓ parallel
fio
vdbench
iozone
iometer
↓ parse all
Comparator (Gemini)
full report + recommendation
~68K
IOPS (sim)
4355
MB/s (sim)
650µs
read lat
<3%
tool variance
Quick start
$ pip install -r requirements.txt
$ export DRY_RUN=true
$ python storage_benchmark_agent.py
API & automation

REST API agent (LangGraph + CrewAI)

Two complete agent implementations — LangGraph and CrewAI — both targeting REST APIs from natural language prompts. Includes retry logic, exponential back-off on 5xx, Gemini free-tier throttling, and a side-by-side comparison runner.

  • LangGraph agent: stateful graph with conditional retry edges and MemorySaver checkpointing
  • CrewAI agent: 4-role crew — researcher, API caller, parser, writer
  • 5s throttle between LLM calls — stays under Gemini's 15 RPM free tier
  • Exponential back-off with jitter on 429 rate limit errors
  • Side-by-side runner compares both frameworks on the same query
  • Demo target: JSONPlaceholder — no auth, runs immediately
LangGraph CrewAI Gemini 1.5 Flash REST API Python 3.9+
LangGraph vs CrewAI
LangGraph — graph-based
Planner node
LLM decides endpoint
↓ tool call
REST executor
HTTP request + retry
↓ 200 OK
Formatter node
plain English answer
CrewAI — role-based
Researcher
API caller
Parser
Writer
Quick start
$ pip install langgraph crewai
$ export GOOGLE_API_KEY="AIza..."
$ python rest_api_agents_starter.py
Kubernetes & cloud

PrestoDB EKS scaling agent

Replaced a fragile Selenium-based browser automation workflow with an AI agent that autonomously scales a PrestoDB cluster on AWS EKS from 0 → 80 Kubernetes pods — supporting large-scale financial data query workloads at a financial institution.

  • Replaces Selenium browser automation with goal-driven agent execution
  • Scales PrestoDB from 0 → 80 EKS pods based on query load signals
  • Monitors pod readiness and coordinator registration before declaring scale complete
  • Fault-tolerant — retries failed pod schedules and handles node pressure
  • Built for a financial institution handling large distributed SQL queries
AWS EKS Kubernetes PrestoDB Python kubectl agentic AI
Scaling flow
Query load signal
CPU / queue depth threshold
Scale decision agent
LLM decides target replica count
↓ kubectl scale
pods 1–20
pods 21–40
pods 41–60
pods 61–80
↓ readiness check
80 pods ready
coordinator registered
0→80
pod scale
0
Selenium deps
NFS & storage testing

Connectathon NFS Agent

Runs NFS Connectathon test suites — basic, general, special, and lock — across NFSv3, NFSv4, and NFSv4.1 or multiple mount points. Parses pass/fail/skip per suite per version, detects regressions, and produces a full comparison report.

  • Covers all 4 suites: basic (b), general (g), special (s), lock (l)
  • Compares across NFSv3, NFSv4, NFSv4.1 and multiple mount points
  • Detects regressions and improvements vs baseline version
  • Explains expected failures (e.g. NFSv4 root squash / setuid)
  • Dry-run mode — full simulation, no NFS server needed
  • Reports saved as JSON + text on every run
LangGraph Gemini 1.5 Flash NFSv3 NFSv4 NFSv4.1 Connectathon Python 3.9+
Agent flow
Prompt
"Run b,g,s,l on server:/export NFSv3 v4 v4.1"
↓ per version
NFSv3
NFSv4
NFSv4.1
mount2
↓ parse each
Compare + report
regressions flagged
4
suites
3+
NFS versions
Quick start
$ export DRY_RUN=true
$ python connectathon_agent.py
OpenStack & cloud testing

OpenStack Tempest Agent

Runs OpenStack Tempest test groups — compute, network, storage, identity, image — across two environment configs (e.g. prod vs staging). Surfaces regressions, improvements, and exact failed test names with a full comparison report.

  • Covers compute, network, storage, identity, image, object groups
  • Compares two configs side by side — prod vs staging, A vs B
  • Lists exact failing test names per group per config
  • Detects new regressions and fixed tests between configs
  • Generates tempest.conf automatically from prompt parameters
  • Dry-run mode — no OpenStack environment needed to test
  • LangGraph Gemini 1.5 Flash OpenStack Tempest Keystone Python 3.9+
    Agent flow
    Prompt
    "Compare prod vs staging, run compute network storage"
    ↓ per config
    prod
    staging
    ↓ parse each
    Compare + report
    regressions + failed tests
    6
    test groups
    2+
    configs
    Quick start
    $ export DRY_RUN=true
    $ python tempest_agent.py