Skip to main content
MKMKConsultantGroup

Journal

Ideas from the desk

Notes on shipping production AI: strategy, architecture, and field stories from logistics to government.

Filter by topic

Fictional threshold comparison for 1,000 shipment messages containing 100 urgent cases. Thresholds 0.90, 0.70 and 0.40 flag 40, 100 and 220 cases, missing 64, 30 and 10 urgent cases. At four minutes per review, only the 160-minute queue fits the illustrative 180-minute capacity.
AI / ML6 min read

AI confidence thresholds: size the review queue

A freight exception example compares missed urgent cases, false alarms and review minutes. Choose a threshold the operation can support, then test it on unseen work.

MK

Muntaz Kaleem

MK Consultant Group

Latest

24 essays

Fictional shipment event trace: version 18 arrives, version 17 arrives late, then event 18 repeats. A durable inbox retains the late event and deduplicates the repeat while the projection stays at version 18. A later agent proposal based on version 19 is held when current state reaches version 20.
System Design6 min read

AI shipment webhooks: handle late and duplicate events

Trace a late shipment update through an event inbox, versioned state and a stale agent proposal. Prevent arrival order from becoming an accidental business rule.

Muntaz Kaleem·Oct 21, 2026

Illustrative prompt-cache comparison with a 6,000-token stable prefix and 1,000 fresh tokens per request. Four requests ten minutes apart create four five-minute writes or one one-hour write and three reads; example input costs are USD 0.0680 and USD 0.0338.
System Design5 min read

Claude prompt caching: measure reuse before choosing a TTL

A worked API cost example shows when a five-minute cache helps, when a one-hour cache wins, and how to diagnose misses without freezing live business data.

Muntaz Kaleem·Oct 19, 2026

Fictional invoice INV-204 has USD 200 in line items, a USD 10 discount, USD 19 tax and USD 15 freight, producing USD 224. Extracted JSON omits the discount and returns USD 234. Schema checks pass but reconciliation finds a USD 10 difference and holds posting for review.
Process6 min read

AI invoice extraction: reconcile line items before posting

A worked invoice shows why valid JSON and matching totals can still hide an extraction error, and how source evidence, decimal checks and review queues help catch it.

Muntaz Kaleem·Oct 19, 2026

Illustrative freight pricing: USD 1,150 cost produces a USD 1,437.50 price at 20 percent margin; 20 percent markup gives USD 1,380. Read source facts, run an approved calculation rule, and hold incomplete quotes.
Logistics6 min read

AI pricing agents: keep the margin calculation in code

A worked freight quote explains margin versus markup, decimal rounding, missing inputs and how programmatic tool calling fits a pricing workflow.

Muntaz Kaleem·Oct 14, 2026

Permission-aware retrieval separates tenant A's allowed documents from tenant B's restricted documents before model context. A revocation timeline shows why stale ACL metadata and cached answers need an explicit policy.
System Design7 min read

RAG document permissions: test revocation and cached answers

A worked enterprise search example explains security filters, permission synchronization, chunk metadata and the caches that can leak restricted answers.

Muntaz Kaleem·Oct 14, 2026

Muse Wonder evaluation map: select Model API or Muse Code, prepare a freight proposal, check current records and authority, then review or execute a permitted action.
AI / ML7 min read

Meta Muse Wonder: what business teams should test

Meta's new Muse Wonder models explained: access, memory, tool use, and a worked freight example for evaluating business agents before rollout.

Muntaz Kaleem·Oct 12, 2026

Illustrative 120-record batch reconciliation: 112 provider successes and eight unresolved results become 102 ready records, seven needing review, three stale proposals and eight unresolved records.
System Design6 min read

Batch AI processing: reconcile every business record

A worked document-processing batch shows how to match unordered results, separate model success from usable output, and prevent stale updates.

Muntaz Kaleem·Oct 12, 2026

Three responsibilities in a CRM assistant: a skill provides the procedure, MCP supplies current records, and application code validates a proposed handoff before any write.
AI / ML5 min read

Agent Skills vs MCP: who owns the business workflow?

Skills package a procedure. MCP connects tools and data. A worked CRM example shows how to combine them, handle conflicting records and evaluate the handoff.

Muntaz Kaleem·Oct 7, 2026

Overnight agent handoff: save a case checkpoint, refresh the shipment record on resume, compare changes and either continue, revise the proposal or assign an operator.
System Design6 min read

Long-running AI agents: resume with current business state

A saved conversation can preserve an old decision. Learn what an agent handoff should retain, what it must refresh and how to test a workflow that pauses overnight.

Muntaz Kaleem·Oct 7, 2026

Choose the model. Keep the policy.. Laya; Jev; OpenAI Decisions; Normalize the decision; Check permission + uncertainty; Execute + inspect.
AI / ML8 min read

AI classifiers for business: Laya, Jev and OpenAI Decisions

A practical guide to AI classification: compare Laya, Jev and OpenAI Decisions, follow a freight request, and evaluate confidence, errors and review workload.

Muntaz Kaleem·Oct 1, 2026

Order recovery flow: approve, submit, timeout, reconcile, then confirm, safely retry, or hold for investigation.
System Design3 min read

Your AI agent timed out. Did it already do the job?

A timeout can hide a completed action. How to design AI workflows that recover without creating duplicate orders or leaving operators guessing.

Muntaz Kaleem·Oct 1, 2026

Invoice security flow: untrusted document, read-only extraction, validation, scoped proposal, independent permission check, execution or review.
AI / ML3 min read

The invoice your AI reads should not control what it does

Document-reading agents need a boundary between incoming text and permission to act. A practical review for teams connecting AI to business systems.

Muntaz Kaleem·Oct 1, 2026

Build the operating advantage. Standard tools; Vendor interface; Relevant business inputs; Encode domain judgment; Fit the workflow; Learn from outcomes.
Strategy2 min read

Why SaaS can't be your competitive moat

If your competitors can buy the same seat, they will. The advantage isn't the model. It's the system wrapped around your operations.

Muntaz Kaleem·Jun 12, 2026

Build the evidence path. Version the sources; Retrieve candidates; Rerank the evidence; Compose the answer; Evaluate the response; Release + monitor.
AI / ML2 min read

RAG is not a product. It's a subsystem.

Hybrid retrieval, reranking, evaluation, and guardrails. The boring layers that decide whether your 'AI app' survives contact with production.

Muntaz Kaleem·May 21, 2026

Carry consent through the workflow. Record the lead source; Check permitted contact; Qualify the reply; Prepare the appointment; Confirm the commitment; Keep follow-up consistent.
Insurance2 min read

Insurance CRM without the VA tax

How omnichannel AI qualification replaced expensive, inconsistent human follow-up, and kept CCPA in the loop.

Muntaz Kaleem·May 8, 2026

Turn fleet signals into feasible action. Collect fleet signals; Investigate the shortfall; Prepare the evidence; Propose an intervention; Check constraints; Verify the outcome.
Energy1 min read

Solar IPP revenue protection in the real world

Curtailment detection, BESS optimization, and PREPA-grade compliance. Monitoring that pays for itself.

Muntaz Kaleem·Apr 28, 2026

Keep eligibility tied to its evidence. Validate the inputs; Select current rules; Prepare the triage result; Review the candidate; Record the decision; Recheck after changes.
Mortgage1 min read

Mortgage underwriting without the cheat sheets

Public filings, intelligent outreach, and real-time eligibility across 32+ state licenses. Triage that agents can trust.

Muntaz Kaleem·Apr 14, 2026

Expiry is part of the transfer decision. Identify the batch; Check stock conditions; Assess receiving need; Propose a feasible transfer; Authorize the movement; Verify the receipt.
Healthcare1 min read

Pharma FEFO at national scale

Near-expiry redistribution, cold chain, and GS1 verification across 400+ facilities. AI that moves medicine, not slides.

Muntaz Kaleem·Mar 30, 2026

Design the service and its handoff. Accessible intake; Check documents; Prepare the service case; Route to the service owner; Resolve or escalate; Record completion.
Government1 min read

Government AI that citizens can actually use

Chatbots, document processing, lab interconnection, and hospital diagnostics. Resident AI partnership, not a pilot graveyard.

Muntaz Kaleem·Mar 12, 2026

Improve judgment. Preserve authority.. Prepare the proposal; Measure agreement; Define limited autonomy; Check the release policy; Review commitments; Record the action.
System Design2 min read

Trust tiers for agentic systems

Proposal-only by default. Autonomy earned from outcomes. External sends and prices stay gated forever.

Muntaz Kaleem·Feb 26, 2026

Build with the people who run it. Observe the desk; Choose a bounded problem; Agree the working contract; Build in the actual workflow; Review with operators; Improve from outcomes.
Process1 min read

Embedded engineering, not slideware

We don't hand you a login. We sit with your ops team, map workflows, and ship in weeks. Then the moat compounds.

Muntaz Kaleem·Feb 10, 2026

Keep the workflow. Replace the worker.. Business workflow; Adapter contract; Evaluation cases; Candidate providers; Release the change; Monitor real outcomes.
AI / ML1 min read

Model-agnostic by design

When the next model drops, swap the layer. Keep the business logic. Architecture beats duct tape.

Muntaz Kaleem·Jan 22, 2026

Make the release evidence visible. Define the task; Build the test set; Run the actual system; Review failure patterns; Decide against thresholds; Monitor after release.
AI / ML1 min read

Evaluation before enthusiasm

If you can't measure hallucination, latency, and task success, you can't govern an agent. Full stop.

Muntaz Kaleem·Jan 8, 2026