IQ IqraLearning

Infrastructure & AI

AI Capacity Management Platform

Capacity answers grounded in tools and live pricing—not chat guesses.

Atif Jaffery · AI Architect — modular prompts, tool contracts, anti-hallucination stack

1 agent · 17 tools · numbers only from tools

01The situation

In 60 seconds

Business problem

Infrastructure leaders need to know whether capacity holds, and for how long, across compute, storage and network. Spreadsheets and chat guesses produce false alarms and invented numbers.

Unacceptable failure

A capacity or cost figure the model made up.

Design move

One agent, many tools. Every figure comes from a registered tool or API, and code injects the mandatory preflight tools if the model skips them.

What it shows

Answers cite the tool endpoints their numbers came from.

The problem

Leaders ask one question: are we okay, and for how long? Answering it means reconciling VMware, storage, SAN and network metrics with growth and pricing. Chat without tools invents the numbers, and short spikes trigger false emergencies.

Who feels it

  • Capacity planners
  • Infrastructure operations
  • Migration and assessment stakeholders

What’s at stake

  • Spreadsheet planning that lags reality.
  • False capacity emergencies from short spikes.
  • Cloud-cost guesses with no pricing source.

Constraints

  • Figures come from tools and APIs, never model memory.
  • The model provider is configurable: LM Studio, OpenAI, Azure, Bedrock or Anthropic.
  • Live AWS pricing with a one-hour cache and a static fallback.

Failures this system must never produce

  1. F1

    An invented metric

    A utilization, growth or cost figure that didn’t come from a tool.

  2. F2

    A recommendation without the checks

    Answering “do we need more?” before pulling the verdict and forecast.

  3. F3

    A spike treated as a trend

    A short spike must not trigger a capacity emergency.

02The approach

How AI solves it

Each failure gets one mechanism, and each mechanism has proof further down this page.

Failure Mechanism The model’s role Proof
F1
Mechanism

Seventeen registered tools return forecasts, verdicts, growth drivers, rightsizing and market pricing, each with a source endpoint. An import-time parity check fails fast if a tool is missing.

The model’s role

The model chooses tools and narrates. It doesn’t supply numbers.

Explorer →
F2
Mechanism

Recommendation questions require mandatory preflight tools. If the model skips them, code injects them before it answers.

The model’s role

None at preflight; it’s enforced in code.

Explorer →
F3
Mechanism

The capacity verdict is reconciled from canonical metrics in code, and business-hours analysis separates working load from spikes.

The model’s role

The model explains the verdict; code computes it.

Guardrails →

Where the model works, and where it doesn’t

Model-assisted

  • Capacity Intelligence AgentModular prompts assembled per intent. Picks tools and writes the answer.

Rule-bound or human

  • CollectorsVMware, storage, network and migration metrics. No model.
  • 17 FastAPI toolsForecasts, verdicts, growth drivers, rightsizing, pricing.
  • Mandatory preflightInjected in code when the model skips it.
  • Capacity verdictReconciled from canonical metrics in code.
  • AWS pricing clientboto3 with a one-hour cache and a static fallback.
  • PlannerActs on the recommendation.

Agents and who holds authority

One agent writes the answer. Code owns every number and forces the checks before any recommendation.

Click any card for its contract: inputs, outputs, decision authority, and guardrails.

LLM model-assisted Code deterministic Human person in the loop
Measures

Collection

Metrics and prices arrive without any model in the path.

Computes

Tools

Every figure the agent can quote comes from here.

Explains

Agent

Chooses tools and narrates the result.

Enforces · Decides

Authority

Code forces the checks; a person acts on the answer.

Guardrails & trust

  • Figures must come from tools/APIs—not model memory.
  • Import-time tool-parity assert and tests fail fast on missing tools.
  • Mandatory preflight for recommendation questions.
  • Prompt bans code/shell exfiltration; re-verify before reversing a verdict.

03See it work

See it run

A 30-second walkthrough of one capacity question, and a live demo on sample metrics.

Live demo · cloud

Coming soon

Ask a capacity question and check the sources

  • Ask whether a sample cluster has headroom
  • Watch the preflight tools run before the answer
  • Open the source endpoint behind each figure
  • Compare a migration option with live AWS pricing
  • Export the PDF report
Open live demo · soon Book a walkthrough
  1. 0:00–0:06 Ask

    The planner asks whether a cluster has headroom.

  2. 0:06–0:14 Preflight

    Code injects the verdict and forecast tools.

  3. 0:14–0:22 Answer

    The agent answers with a cited source for every figure.

  4. 0:22–0:30 Price

    A migration option priced with live AWS data.

Interactive architecture

Walk a real request through agents, services, stores, and human gates. Click any node for its contract.

04The result

What changes for infrastructure planning

A longer planning horizon

Fewer false spike alarms (F3).

Migration conversations backed by cited metrics

Every figure traces to a tool endpoint (F1).

Cost answers with a pricing source

Live AWS pricing, with explicit fallback behavior.

How I check it behaves

Each check has a definition and a target. Results appear only from real runs, with the set size and date.

Check What it counts Target Last run
Untraced figures F1 Numbers in answers without a tool source 0 Run pending
Tool parity F1 Registered tools matching the router (test_chat_tool_parity) Pass Run pending
Preflight coverage F2 Recommendation answers where preflight ran All Run pending
Spike false alarms F3 Emergencies raised from short spikes on a replayed set 0 Run pending

05Looking back

Architecture decisions

  • Code-enforced preflight F2

    So the checks don’t depend on the model choosing to call them.

    Cost acceptedSimple questions take longer to answer.

  • Modular prompts per intent

    So migration and capacity answers stay focused.

    Cost acceptedMore prompt modules to keep in sync.

  • Collectors separate from the language agent F1

    So metrics flow whether or not the model is involved.

    Cost acceptedTwo systems to operate.

Known limits and next step

Scope · disclosed

Live pricing is built for AWS. Pricing depth for other clouds isn’t confirmed.

Treat non-AWS cost comparisons as indicative until pricing parity is confirmed.

Next step

  1. Confirm pricing depth for each provider before quoting it.
  2. Update the README’s tool count: it says 9, and the router registers 17.

Project demo