← Case studies
Industry / 03 LLM + optimization Decision science Supply chain

Supply Sense: ask the optimizer in plain English.

A plain-English layer over supply-chain optimization. A planner asks a question; the system writes the optimization, runs the solver, and explains the trade-off back in words, without ever moving proprietary data.

My role
AI strategy & delivery lead
Built at
IBM Consulting · applied GenAI
Methods
LLM · in-context learning · MIP
Recognition
IBM Technical Premier Award
Status
Reusable IBM asset

Demo note. The interactive demo runs on synthetic data and the client is not named. This is a sanitized recreation of the delivered system, built to show the approach without exposing confidential details.

Interactive · ask the scenario manager, watch the network re-optimize
Synthetic data
Scenario Manager Chatbot
Ask a what-if, or pick a scenario above…
Solver · optimal · baseline plan

A working sketch of the real thing: a planner asks the scenario manager a what-if in plain English; the system rewrites the underlying mixed-integer optimization, re-solves it, and answers with the trade-off, recommended next moves, a projected trend, and the intermediate solver code for anyone who wants to audit the reasoning. The live network on the right re-routes to each new optimum. The production system did this with in-context learning, so proprietary data never left the enterprise.

A day in the life

How a planner actually uses it, notification to action.

Composite persona: a hub-operations planner whose job is to keep daily capacity safe across a 26-hub network, no overflows, no wasted volume.

01 · Notification
The optimizer flags a risk
The system recommends temporarily blocking in-gates at one hub: forecasted volume will breach its safe capacity ceiling later in the week.
02 · Formulate
Probe it in plain English
The planner digs into root cause, dwell time, lift capacity, the inbound pipeline, and runs hypotheticals: "what if we hold units at the neighboring hub?" Every scenario returns KPI deltas, the reasoning, and the generated solver code.
03 · Validate
See before / after
Impact is visualized on the hub-efficiency trend before anything is committed, and hub leads confirm the plan is feasible on the ground (enough manpower for the added lift).
04 · Act
Commit, and leave a trail
Allocations go into the system and gate restrictions are placed. The scenario, its parameter changes and the reasoning are logged to history, shareable with the whole team or kept private.
05 · Monitor
Watch, and adjust
Capacity is tracked against the plan. If the strategy drifts, saved scenarios are revisited and rebuilt, and planner feedback tunes what the system recommends next time.
Gen-AI capabilities Automated code generation Dynamic visualization Human-in-the-loop Contextual reasoning
01 · The problem

Optimization solves the problem. Then no one trusts the answer.

Modern supply chains demand sophisticated optimization, but the output is a black box. A planner is handed a decision with no sense of why, and no safe way to ask "what if the supplier's lead time doubles?" The math is sound; the trust isn't. And the obvious fix, pipe everything into a general LLM, means leaking proprietary network data.

02 · The system

Supply Sense: a translator between people and solvers.

I built Supply Sense, a system that pairs a large language model with combinatorial optimization. It turns a plain-text question into a real answer about an optimization outcome, and lets a planner reshape the problem (add or drop a constraint, run a what-if) in conversation, while the proprietary data stays put.

It was proven on a live workload: server placement inside IBM's own cloud supply chain.

"A 90% accuracy rate for in-distribution queries, improving inventory turnover, reducing development time, and shortening order cycles."

from an independent reviewer's assessment
03 · How it works

Question → optimization → solver → plain-English answer.

A small set of agents runs the round trip. One translates the question into optimization code; the solver processes it; the result comes back as business outcomes in plain language. Underneath is a mixed-integer model maximizing net revenue across supplier cost, inventory and capacity, subject to constraints like supplier limits, lead times and safety stock.

  • In-context learning, not fine-tuning, the model adapts to the domain from a handful of examples, which is cheaper and far safer.
  • Live what-if analysis, constraints can be added or removed on the fly and the problem re-solved, exactly like the demo above.
  • Privacy by design, it reasons over the structure of the problem, never the raw enterprise data.
04 · Leading it

I led the team and built the core myself.

I led a small team of operations-research scientists and ML engineers and worked directly with the business partners who'd use it. The system drew attention from executives across industries, the CTO of the world's largest freight railroad, the supply-chain VP of the world's biggest food manufacturer, and leaders at the largest U.S. drug distributor, and is now recognized as a reusable IBM asset.

05 · Impact

From an idea to an IBM asset.

90%
accuracy on in-distribution planner queries
~$1M
revenue generated for IBM from the framework
4
industries in scope, air, rail, retail, CPG
Top 0.01%
IBM Technical Premier Award, bestowed by the CEO

For this work I received the IBM Technical Premier Award, the firm's most prestigious technical honor, reserved for the top 0.01% of global contributors across research, software and consulting.

06 · Talks & recognition

Where this work was shared.

NYC
LLMs Meet the Solver: Conversational Optimization
IBM Innovation Studio, executive workshop
Workshop · Host
Conference
Interpretable Supply-Chain Optimization with Language Models
INFORMS Annual Meeting, Practice track
Conference
Award
IBM Technical Premier Award
Honored by the IBM CEO for technical contribution & client impact
Award
Next · Industry / 04
AI Demand Planning