Edmonton, Alberta Est. 2024

The Limbachia Chronicle

"All the Code That's Fit to Ship"



ML & Full-Stack Engineer
Ships Safety-Critical Systems

Soham Limbachia builds production machine learning and the PERN-stack platforms that operationalize it —
React, Node, Postgres, AWS on the front; multivariate ML pipelines and live MLOps on the back.

I'm a Computer Engineering student (UAlberta, Software Option Co-op, graduating June 2027) who builds end-to-end systems: the machine learning that makes decisions, and the full-stack platforms that put those decisions in front of the humans who act on them.

This year at BossPac Engineering & Technology, I designed and deployed a production multivariate anomaly detection pipeline for broken-rail prediction — the leading cause of mainline freight derailments. The pipeline produces a second-opinion vote on threshold-based alarms and has cut false-positive alerts by ~52% in live operation across Class 1 freight subdivisions in North America. I own the MLOps layer — rebaselining, drift detection, per-subdivision scalers, and precision/recall evaluation against confirmed outcomes.

I also built the full-stack platform that operationalizes it. BRIM runs on the PERN stack (PostgreSQL, Express, React, Node.js) deployed on AWS with Redis session management, Cypress and Jest test suites, and AWS Cloudscape components — delivering real-time geospatial alerts, configurable condition-based dashboards, and field-deployment tooling to 10–50 rail operators across 7,650+ miles of Class 1 freight track. The ML is only useful because the platform makes it legible to the humans making safety calls.

Before BossPac, I worked in Process Automation & Controls at Suncor Energy's Base Plant, a ~350 kbpd Alberta oil sands operation, where I developed an input signal conditioning algorithm for a hydrogen flare controller that enabled Advanced Process Control engineers to commission a Model Predictive Control strategy now delivering ~$13K/month in sustained operational savings. I also built a site-wide PLC diagnostic web app adopted by ~10 daily engineers and technicians, cutting fault diagnosis time from 30+ minutes to under 10. On the side, I red-team frontier multimodal AI models as an RLHF contractor at Outlier AI, focused on reasoning-chain failures, hallucination patterns, and tool-use errors across 300+ technical tasks.

I'm looking for teams where both sides of the stack matter — where the model has to work in the field and the interface has to work for the user. Industrial AI, autonomous systems, robotics, safety-critical infrastructure, and applied AI at companies that care about the gap between a model in a notebook and a model in production. I'm equally happy owning general full-stack or backend work at companies building things that matter; the common thread is real-world systems where correctness, reliability, and human trust are non-negotiable.

ML Engineering / Production

Multivariate Anomaly Detection for Broken-Rail Prediction

Production multivariate anomaly detection pipeline deployed across Class 1 freight subdivisions in North America — the leading cause of mainline freight derailments. Produces a second-opinion vote on threshold-based alarms; cut false-positive alerts by ~52% in live operation. Full MLOps layer: rebaselining, drift detection, per-subdivision scaler management, and precision/recall evaluation against confirmed break outcomes via weekly alert-outcome audits.

Full-Stack / Production

BRIM — Rail Integrity Management Platform (PERN on AWS)

Production full-stack rail monitoring platform on the PERN stack (PostgreSQL, Express, React, Node.js) deployed to AWS with Redis session management, AWS Cloudscape components, and Cypress + Jest integration suites. Serves 10–50 rail operators across 7,650+ miles of Class 1 freight track — real-time geospatial alerts, configurable dashboards, and field-deployment tooling. Operationalizes the anomaly detection pipeline for the humans making dispatch decisions.

View BRIM →

Shipping Summer 2026

Automation Part Picker — LLM-Agent Control System Designer

Agentic system that takes plant-level requirements — "3 mixing tanks, Class 1 Div 2 hazardous area, pressure + temperature monitoring" — and produces validated PLC, I/O, sensor, and network bills of materials with hazardous-area, power-budget, and topology checks. Phase 2 adds a process-optimization layer recommending AI vision systems and advanced sensors where they enable new control strategies. Python, FastAPI, Postgres + pgvector, LLM tool-use, evaluation harness with expert-validated design scenarios.

Follow on GitHub →