Selected work
Builds that met real traffic
These anonymised case studies describe past client engagements by MindForge AI, a Toronto applied-AI studio. Metrics are illustrative results from those projects — not promises of what your organization will achieve. Every build included human-in-the-loop oversight and responsible AI guardrails.
Dispatch policy assistant with RAG
The problem: Dispatchers at a Canadian logistics scale-up spent minutes hunting through PDF routing policies every time an exception arose — a support inbox drowning at 2 a.m. during peak season.
Our approach: A discovery sprint mapped 400+ policy pages into a retrieval-augmented generation knowledge base. We built an AI assistant with prompt engineering tuned to logistics vocabulary, guardrail design for out-of-scope queries and mandatory human-in-the-loop escalation when confidence dropped below threshold.
What shipped: Production deployment with MLOps monitoring, model evaluation benchmarks and API integration to their dispatch console. Agent workflow routed complex cases to senior dispatchers automatically.
What stayed human: Final routing decisions on high-value loads, policy exceptions and anything flagged by guardrails. In past deployment, average policy lookup time dropped substantially — your results will depend on data quality and adoption.
Patient intake automation with review gates
An Ontario healthcare provider needed workflow automation for intake forms — not a replacement for clinical staff. We prototyped an NLP extraction pipeline that pulled structured data from scanned forms, validated fields against business rules and pushed clean records via API integration to their EHR. Every flagged field required human review before submission. PIPEDA-compliant data privacy governed storage and retention. The proof of concept took four weeks; production deployment added MLOps alerting and model evaluation for extraction accuracy. Clinical judgment remained entirely with licensed staff.
Inventory anomaly detection
A national retailer had a spreadsheet nobody trusted for regional stock alerts. We built a machine learning model with data pipelines ingesting POS and warehouse feeds, plus a dashboard for operations managers. Computer vision was not needed — classical ML with careful feature engineering solved the problem faster than a generative-AI approach would have. Model evaluation against historical shrink data guided threshold tuning. The system runs in production with weekly eval runs under a retainer engagement. Measurable outcomes improved in past work; we do not guarantee cost savings or ROI for future clients.
Internal compliance Q&A assistant
A Toronto enterprise needed an AI assistant for internal compliance questions — high stakes, zero tolerance for fabricated answers. We deployed a retrieval system over approved policy documents only, with guardrails blocking generative responses when no source chunk matched. Fine-tuning was deliberately avoided in favour of strict RAG with citation requirements. Human-in-the-loop review applied to every answer before it could be shared outside the compliance team during the pilot phase. Production deployment expanded access with logging and responsible AI audit trails. This is the kind of custom build our Toronto AI studio specialises in: conservative engineering for regulated contexts.
Case study metrics reflect past client work and are not promises of future performance. AI systems can err; we design for oversight, not autonomy.