Persistent Systems integrated yield-prediction copilots into live blast-furnace telemetry at Tata Steel’s Jamshedpur works this month, according to people familiar with the deployment. The software sits on top of historian data from burden control, hot-metal temperatures, and off-gas chemistry, surfacing recommendations operators can accept or reject from hardened consoles—not from consumer chat interfaces on the plant floor.
What the copilot actually does
Blast furnaces are continuous reactors; small shifts in coke rate or blast moisture ripple into hot-metal silicon levels hours later. Veteran operators read trends from experience; the copilot trains on years of tagged episodes where teams documented what worked before a tap. Persistent’s team built feature stores that align sensor clocks—no trivial task when legacy PLCs stamp timestamps differently than newer IoT gateways on the same aisle.
Outputs are bounded: the model proposes adjustments within envelopes Tata Steel’s metallurgists pre-approved. If a recommendation would cross a safety interlock, the UI grays it out and logs the attempt for audit. There is no autonomous closed loop; plant managers insisted on that after reviewing MeitY-adjacent guidance on industrial AI and after internal unions asked who signs off when silicon drifts.
Deployment architecture
Inference runs on an on-prem cluster air-gapped from public cloud APIs. Persistent shipped containerized services that pull from the historian through OPC bridges, score every few minutes, and write explanations alongside scores so engineers see which sensors drove a nudge. When the cluster fails, furnaces keep running on manual mode—the copilot is advisory, not load-bearing.
Tata Steel’s IT security team ran red-team exercises on the inference API to ensure a compromised laptop in the office VLAN cannot push set points. Role-based access mirrors the plant’s existing shift handover badges; contractors lose tokens automatically when their passes expire at the gate.
Evaluation metrics the board watches
Leadership tracks hot-metal chemistry variance, specific consumption of coke, and the rate at which operators override suggestions. Early weeks show more overrides on swing shifts when junior crews staff the pulpit—exactly where trainers hoped the copilot would act like a silent senior. Persistent and Tata Steel agreed not to publish yield percentages yet, citing competitive sensitivity, but internal dashboards compare tap-to-tap stability against the same monsoon quarter last year.
Ministry of Steel officials promoting decarbonization want better fuel efficiency without capital rebuilds; software layered on existing furnaces fits that narrative. NITI Aayog’s industry workshops have cited steel as a sector where India can export operational AI if plants prove governance first.
Data governance and unions
Collective bargaining units asked whether shift logs used for training anonymize crew identifiers. Tata Steel said yes, and that override buttons exist so no operator is penalized for ignoring a bad suggestion. Persistent’s contracts include model retraining windows where union safety reps can flag episodes that must never enter training sets—near-misses, for example.
Regulators have not issued blast-furnace-specific AI rules, but DPIIT’s responsible-AI frameworks echo in vendor RFPs: traceability, human oversight, and incident reporting. Tata Steel extended its internal digital ethics checklist to cover copilot releases with the same rigor as customer data platforms.
Limits and hallucination risk
Large-language-model hype does not match this build. The copilot’s natural-language layer explains recommendations in Hindi and English, but numeric cores are gradient-boosted and physics-informed models validated against lab assays—not general chatbots guessing at chemistry. Persistent engineers said they refused a demo that would let operators free-type questions about the furnace; unstructured prompts are disabled in production.
When sensors drift out of calibration, the system downgrades confidence and flashes maintenance tickets. That behavior matters during Jharkhand dust season when particulate fouls air intakes and models trained on clean monsoon data would otherwise overfit.
What comes next
If Jamshedpur stabilizes metrics through Diwali peak demand, Tata Steel may extend similar copilots to sinter plants and coke ovens, each with separate safety envelopes. Persistent hopes the engagement becomes a reference for other metal majors negotiating on-prem AI without sending raw telemetry abroad.
For now, the story is narrow: a services firm and a steel giant agreed that yield prediction belongs beside the pulpit, not in a slide deck—and that human veto power is a feature, not a bug.
Instrumentation gaps the team closed
Before models shipped, instrument technicians replaced failing thermocouples on top-gas lines whose stale readings had been confusing older advisory tools. Persistent’s data engineers tagged those weeks as “do not train” so the copilot would not learn from garbage inputs. The cleanup delayed go-live by ten days but cut false positives in shadow mode.
Tata Steel also linked the copilot to its maintenance work-order system: when confidence drops because a sensor flatlines, a ticket opens automatically with the asset ID operators already use in the field. That integration is mundane IT work, but it is what separates a pilot from something crews trust on a hot afternoon before tap.
