Ingestion, storage, latency and failure modes decide whether a model is useful in practice. Everything on this page was built inside one professional role, for external industrial clients.
Organisation
Hi-Tech System & Services
Period
Dec 2024 – Oct 2025
Engagements
07
E.01The Role
IndustryProfessional engagement
Hi-Tech System & Services Ltd.
Software Engineering & Machine Learning Intern
R&D Division · Kolkata, India
Period
Dec 2024 – Oct 2025
Deployed for
IISCO Steel Plant · JUSCO · Yara International · NTPC
Engagements
07
Eleven months in an R&D division building systems installed on live industrial sites — a steel plant, a utility operator, an ammonia terminal and power boilers — where the cost of a missed detection is measured in safety and production, not validation loss.
Torpedo ladles carrying molten steel can develop leaks in transit. Detection relied on human observation across a continuous operation, so an incident could progress before anyone intervened — with direct safety and production consequences.
System
A real-time computer vision system reading thermal camera feeds on the transfer route, classifying thermal signatures continuously and raising an alert the moment a leak signature appears.
Technical approach
YOLOv10 object detection was applied to thermal imaging to isolate the heat signature of escaping molten steel against the ambient thermal background. The detector runs against the live feed under 24/7 operating conditions rather than batch evaluation.
Result
Undetected incidents reduced by 95%, preventing ₹12L/month in production loss across 24/7 operations.
Thermal fieldSCHEMATIC
leak signature
Detector isolates the escaping-steel heat signature against the ambient thermal background.
95%
reduction in undetected incidents
₹12L/month
production loss prevented
24/7
continuous operation
System screens03 · Hover to reveal
Pl. 01 — Monitoring dashboardLive event view
Pl. 02 — Thermal detectionCamera feed
Pl. 03 — Temperature matrixPer-cell °C
[‡]Captures of the deployed Torpedo Monitoring System, from the Hi-Tech System & Services project archive.
E.03Further Engagements
Six more systems,same role.
IndustryDelivered for IISCO Steel Plant, JUSCO, Yara International, NTPC under the R&D Division role at Hi-Tech System & Services Ltd..
02
RFID Inventory Management
IISCO Steel Plant
Backend system
PostgreSQLREST APIRFID
Problem
Component movement was recorded manually, so inventory state was persistently out of date and staff were absorbing a steady rate of transcription errors.
System
An RFID-based inventory management system with a PostgreSQL backend and a REST API serving the tracking interface used across daily operations.
Approach
RFID reads are normalised into event records against a relational schema built for component-level history, then exposed through a REST API so every downstream consumer reads one authoritative state instead of keeping its own.
Result
Served 200+ daily users and tracked 5,000+ components with 100% accuracy, eliminating 8–10 manual errors per week.
5,000+
components tracked
200+
daily users
100%
tracking accuracy
03
Incident Response Pipeline
IISCO Steel Plant
Operational tooling
Event-Driven ArchitectureAutomated Alerting
Problem
Detecting an incident is only half the problem. Response time was dominated by manual escalation, so the interval between detection and action stayed long enough to matter.
System
An event-driven incident response pipeline with automated alerting and a standardised escalation workflow.
Approach
Incidents are emitted as events and routed automatically to the responsible party, replacing ad-hoc escalation with a standardised workflow so response no longer depends on who happens to notice.
Result
Mean-time-to-respond reduced by 40%, from 25 minutes to 15 minutes.
40%
faster mean-time-to-respond
25→15 min
response interval
04
Camera Calibration System
JUSCO
Embedded systems
FFT Signal ProcessingPythonClosed-Loop Control
Problem
Pan-tilt cameras drift out of alignment over time. Correcting them by hand meant a daily manual pass across every installation, and the equipment was unavailable while it happened.
System
An automated drift-correction system that measures misalignment from the captured signal and closes the loop without an operator.
Approach
An FFT-based signal processing algorithm extracts the drift component from the captured signal, feeding a closed-loop control system written in Python with real-time feedback that converges on the corrected position.
Result
Positional accuracy within ±0.5° — a 4× improvement over the manual calibration baseline — reducing equipment downtime by 40% across 12 CCTV installations and eliminating daily manual adjustments.
±0.5°
positional accuracy
4×
improvement over manual baseline
40%
downtime reduction
05
Ammonia Tanker Level Measurement
Yara International
Ammonia terminal, West Bengal
FLIR LWIRPyTorchOpenCVModbus/TCPPLC
Problem
Loading road tankers with pressurised anhydrous ammonia risks catastrophic overfill. Contact sensors require penetrating the shell and drift out of calibration, so the measurement you depend on is the one most likely to fail.
System
A non-contact thermal system that reads the liquid level from outside the tanker and drives an automatic hard-stop at the fill setpoint.
Approach
A FLIR long-wave infrared camera images the tanker shell; a convolutional network locates the vessel and estimates the liquid–gas meniscus through heavy vapour, with OpenCV tracking the boundary frame to frame. At the 80.77% setpoint the system triggers a PLC solenoid hard-stop over Modbus/TCP, and every session is logged locally.
Result
External thermal level tracking with an automatic hard-stop at the fill setpoint, removing the need to penetrate the tanker shell.
80.77%
automatic hard-stop setpoint
Non-contact
no shell penetration
System screens03 · Hover to reveal
Pl. 01 — Tanker detectionFLIR LWIR feed
Pl. 02 — Level trackingFill in progress
Pl. 03 — Operations dashboardSession overview
06
Acoustic Steam Leak Detection
NTPC Limited
Utility power boilers
Airborne Mic SensorsWaveguide TubesPython DSPModbus/TCP
Problem
High-pressure boiler tubes fail from ash erosion, and a small leak widens fast while damaging neighbouring tubes. If it is only caught when pressure drops, the result is an unscheduled shutdown.
System
Continuous online acoustic monitoring that hears a leak forming long before it registers as a pressure loss.
Approach
High-sensitivity airborne microphones in purge-air waveguide tubes capture internal sound pressure. Python DSP isolates the characteristic high-frequency hiss of escaping steam from the broadband rumble of combustion, with a .NET service handling Modbus/TCP alarm signalling.
Result
Tube leaks detected ahead of traditional pressure-drop indicators, converting unplanned outages into scheduled maintenance.
Early
detection before pressure drop
Online
continuous monitoring
System screens02 · Hover to reveal
Pl. 01 — Acoustic trenddB per channel
Pl. 02 — Alarm viewChannel monitor
07
Substation Monitoring
JUSCO
IoT data platform
Event-Driven MicroservicesTime-Series DatabaseRESTful API
Problem
Substation faults are only actionable if they surface immediately. Telemetry arrives continuously, so any batch reporting model reports the fault after it has already cost something.
System
A distributed real-time telemetry system with a web dashboard and threshold-based alerting over live signals.
Approach
Temperature, voltage and load signals are consumed by event-driven microservices and written to a time-series database suited to high-frequency ingestion, with a RESTful API and dashboard layered on top and alert thresholds evaluated against the live stream.
Result
Sustained 500+ sensor signals per minute with proactive fault detection, reducing transformer downtime by 10% in the first quarter.
500+
sensor signals / minute
10%
transformer downtime reduction
Elsewhere
Independent workis recorded separately.
Self-directed systems work and the ISI research collaboration were built outside any employer.
