Industrial Mind & Code · Independent Applied Research

Where should AI sit inside industrial systems?

Placing probabilistic intelligence inside deterministic industrial systems.

Industrial Mind & Code is an independent applied research programme by Siddharth Srinivasan, testing LLMs and agents inside industrial engineering decision environments — supply chains, maintenance systems, production planning, manufacturing workflows, and product-lifecycle systems.

The early result is not “AI beats everything.” It is more interesting: models often fail when placed directly in control of operational decisions, but become useful when constrained inside the right architecture.

I separate a bare LLM (the model on its own) from an agent (an LLM paired with tools and a control loop), and ask where each can sit inside a deterministic industrial system — and where it should be constrained by one.

The bias is practical: compare AI agents against deterministic baselines, measure operational outcomes, and identify where hybrid architectures are more reliable than direct agentic control.

Researcher

Siddharth Srinivasan

Domain

Industrial Engineering × AI

Focus

Agents, control loops, deterministic baselines

Mode

Independent applied research

Core question

Where should probabilistic intelligence sit inside deterministic industrial systems?

Industrial Mind & Code is not asking whether AI agents are impressive in isolation. It is asking where they belong inside real industrial systems.

The early answer: not everywhere, not directly, and not without deterministic constraints.

Emerging Thesis

01

You cannot prompt your way out of the bullwhip effect.

Prompting helped structure behaviour, but it did not make LLM agents outperform simple deterministic replenishment rules. In the intent-classification experiments, a neutral-prompt instruction had zero effect on order variance.

02

More context can make an industrial agent worse.

Additional context can increase reactivity and instability if the agent is placed in the wrong control role. In the hybrid safety-stock experiment, added context increased order variance for two of three models.

03

Perfect intelligence in the wrong position is worse than no intelligence.

Better labels, cleaner intent, or more model capability do not fix a poor control architecture. In the oracle-label study, perfect ground-truth labels produced worse order variance than the formula running alone.

04

Reasoning tokens are not operational intelligence.

More deliberate model reasoning does not automatically translate into better industrial system performance. In the baseline replenishment study, reasoning models generated over 1,000,000 reasoning tokens with no measurable improvement in ordering performance.

05

The question is not whether to use AI, but where to place it.

The strongest results appear when AI is bounded inside a hybrid architecture rather than given direct operational control. In the replenishment experiments, order variance fell below 1.0 only once the AI selected the smoothing parameter and the deterministic formula still placed the order.

The practical design question is therefore not whether AI should replace established industrial methods. It is where probabilistic reasoning can safely improve deterministic systems — and where it should be constrained by them.

Research Tracks

Track 1

Agentic Bullwhip Effect

Experiments testing whether LLM agents amplify or reduce instability in supply-chain replenishment systems.

  • LLM agents vs heuristic baselines
  • Context-window experiments
  • Reasoning-model experiments
  • Oracle label / intent classification
  • AI-in-the-loop smoothing parameters

Track 2

Edge AI for Maintenance

Experiments testing local and edge-deployable AI for maintenance, TPM workflows, sensor interpretation, and vernacular industrial data.

  • FailureSensorIQ
  • TPM / maintenance intelligence
  • Local model behaviour
  • Vernacular normalisation

Track 3

Industrial AI Architecture Patterns

Experiments and essays focused on where AI should sit in industrial control loops and enterprise architectures.

  • Hybrid deterministic/probabilistic systems
  • Semantic layers
  • Control-loop placement
  • Governance and observability implications

Experiment index

Supply chain Control loop LLM agents
The Architecture That Finally Worked: Adaptive Smoothing in Supply Chain Replenishment Published

Question Does AI perform better when it directly controls operational quantities, or when it selects the smoothing parameter inside a deterministic control loop?

Result The first architecture in this program where every AI condition damped order variance (OVAR below 1.0), with the best condition matching and marginally beating the fixed α = 0.3 baseline.

Why it matters Hybrid control architectures may be more promising than fully agentic operational control.

Supply chain Control loop Deterministic baseline
The Ceiling Is in the Formula: Oracle Labels and Structural Incompatibility Published

Question If the replenishment formula is handed perfect, ground-truth labels — no LLM at all — can it finally beat exponential smoothing?

Result No. Oracle labels produced OVAR 1.776, worse than the formula with no AI at 1.753 — and the gap to exponential smoothing is invariant to label quality.

Why it matters Perfect intelligence in the wrong position is worse than no intelligence; the ceiling is architectural, not a model-quality problem.

Supply chain LLM agents Intent classification
Perfect Compliance, Wider Swings: The Five-Label Intent Classifier Published

Question Does confining the AI to a clean five-label classification task keep orders steadier than a standard forecasting formula?

Result No. Every AI condition amplified order swings roughly 6–7× above exponential smoothing — even with perfect label compliance across every run.

Why it matters A reliable interface is not a safe architecture; the fixed label→multiplier table capped what any condition could do.

Supply chain LLM agents Intent classification
The Equaliser Effect: Intent Classification in Supply Chain Replenishment Published

Question Does giving the AI a clean five-label intent classification, instead of a raw number, let it lower order variance?

Result No. Four models from 14B to 120B produced OVAR within a 0.05-unit band, clustering on the formula baseline regardless of size, reasoning capability, or information level.

Why it matters You cannot prompt your way out of the bullwhip effect — model capability and context were discarded at the lookup table.

Supply chain Control loop Hybrid architecture
Hybrid AI Safety Stock Control in Supply Chain Replenishment Published

Question If a deterministic formula executes the orders and the AI only sets the safety-stock multiplier, does the hybrid beat the baseline?

Result No. All four hypotheses rejected — the best AI condition produced roughly four times the baseline’s order variance, and context made two of three models worse.

Why it matters A bounded role is not enough on its own; the AI was still placed on the wrong control surface.

Supply chain LLM agents Deterministic baseline
LLM Agents Against Heuristic Baselines in Supply Chain Replenishment Published

Question Can any LLM configuration beat simple heuristic ordering rules in a serial supply chain?

Result No. Every heuristic outperformed every LLM configuration on both order variance and stockouts.

Why it matters This challenges the assumption that more capable models can be dropped directly into operational decision loops.

Supply chain LLM agents Sovereign models
sarvam-30b in Supply Chain Ordering: A Comparison with GPT OSS 120B Published

Question Does a model trained on Indian data read Indian seasonal demand better than a frontier reference — and can either beat the formula?

Result No measurable difference between the models — and exponential smoothing beat both by 8× on order variance.

Why it matters Model provenance did not change operational behaviour; the deterministic baseline advantage persisted.

Supply chain LLM agents
Context and Model Capability in AI-Driven Supply Chain Ordering Published

Question Does more domain context, or more reasoning capability, reduce how much an LLM amplifies demand swings?

Result Neither reliably helped. Context reduced order variance for the lightweight model and increased it for the reasoning model; every configuration amplified demand swings (directional, n = 5).

Why it matters The first evidence in this program that context interacts with the decision role instead of simply improving it.

Edge AI Fault diagnosis
A Capable Assistant, Not a Decision-Maker: Small Language Models on IBM’s FailureSensorIQ Published

Question Can a model small enough to run locally — on a box beside the machines, no cloud — do real fault diagnosis, and what decides whether its output is usable?

Result The best 4B-class model scored 51.8% on 2,667 benchmark questions — above blind guessing, below the human-expert mean: a capable assistant, not a decision-maker.

Why it matters On the edge, the integration — sampling settings and serving runtime — decides usability as much as the model weights do.

Supply chain Control loop
Cross-Tier Visibility and the Context Penalty: Adaptive Smoothing with Downstream Demand Access In Progress

Question Does giving each tier visibility into downstream demand eliminate the α-inflation context penalty found in the adaptive-smoothing study?

Status In progress — results not yet reported.

TPM Edge AI
Understanding LLM Agentic Capabilities in Total Productive Maintenance (TPM) In Progress

Question Can an LLM do Total Productive Maintenance decision support over fragmented, multilingual shop-floor maintenance records — and does a vernacular-normalisation layer help?

Status In progress — results not yet reported.

Methodology

Industrial Mind & Code evaluates AI agents as components inside operational systems, not as isolated chat interfaces.

  1. Compare against deterministic baselines.
  2. Measure operational outcomes, not just textual quality.
  3. Test model placement inside the architecture.
  4. Separate model capability from control-loop design.
  5. Prefer repeatable simulations and transparent assumptions.
  6. Vet every claim before treating it as a finding.

Infrastructure

Frontier
Azure AI Foundry Azure OpenAI Service
Local
ASUS Ascent GX10 Ollama llama.cpp
Code
Claude Code Codex

Get in touch

LinkedIn  ·  GitHub