Ontario precision manufacturer
Vision model that stopped crying wolf on glare
The problem: A surface-defect detector worked in the lab and failed on the floor — every shift change in lighting triggered false stops. Maintenance ignored alerts; the model became shelfware.
Our approach: Discovery reframed the task from pixel-perfect classification to operator-trustable triage. We rebuilt the labelling guide with floor supervisors, augmented training data for glare and shadow, and shipped an evaluation harness tied to real line footage — not a holdout set from last year.
What we built: Retrained computer vision pipeline, human-in-the-loop override UI, MLOps monitoring on precision/recall by shift, and weekly review cadence with quality leads.
What stayed human: Scrap disposition, line shutdown calls and customer-facing quality claims remained with licensed staff. The model suggested; it never rejected parts autonomously.
Illustrative outcome: False stop rate on two pilot lines dropped by roughly 30% over eight weeks post-launch — measured against pre-build baseline, not a marketing guarantee.
National insurer — claims operations
Policy Q&A with citation discipline
The problem: Adjusters searched twelve PDF versions for the same exclusion clause. A generic chatbot hallucinated confidently; trust collapsed in week one.
Our approach: RAG pipeline over versioned policy corpus with role-based access, chunking tuned for legal cross-references, and retrieval confidence thresholds that route low scores to senior reviewers — not to end users.
What we built: Generative AI assistant integrated into existing workflow tooling, golden-set evaluation harness, prompt versioning, and fine-tuning deferred until retrieval quality justified the cost.
What stayed human: Coverage decisions, client communication and settlement authority remained with licensed adjusters. The assistant assembled context; humans decided.
Illustrative outcome: Median time to locate clause language dropped from ~22 minutes to ~6 minutes in structured UAT samples — adoption varied by desk and document freshness.
Canadian logistics scale-up
Demand sensing model with honest sparsity limits
The problem: Regional managers forecasted replenishment from gut feel because the internal ML model missed every promo spike. Data lived in three systems that disagreed.
Our approach: AI strategy phase named six noisy signals and two authoritative ones. We unified API feeds, trained gradient-boosted models with documented error bounds, and refused to promise accuracy on SKUs with insufficient history — a conversation procurement needed to hear.
What we built: Predictive ML pipeline, evaluation dashboard by region, human override on every automated replenishment suggestion above a dollar threshold, MLOps retraining on a quarterly cadence.
What stayed human: Vendor negotiations, markdown calls and promotional calendar changes stayed with category managers.
Illustrative outcome: Pilot regions reported 10–14% fewer emergency transfers during a 90-day window — past result, not a promise for other networks.
Financial services — Bay Street adjacent
Exception triage assistant for settlement ops
The problem: Analysts reconstructed failed settlement context by copying from five systems into one email thread — forty-five minutes per exception, every afternoon.
Our approach: Retrieval-augmented generation over ticket history, settlement logs and internal policy docs. Prompt engineering tuned for tabular precision; guardrails blocked any output without source citations.
What we built: AI assistant with human-in-the-loop feedback loop, model evaluation on golden-set exceptions, production deployment with rollback after a bad embedding refresh taught us to version corpora properly.
What stayed human: Settlement approval, client communication and regulatory filing remained with licensed operations staff.
Illustrative outcome: Average context assembly time dropped from ~45 minutes to ~11 minutes in user-reported samples during UAT — variance by desk and exception type.
Healthcare admin — Ontario
Referral routing classifier with PIPEDA guardrails
The problem: Referral forms arrived as unstructured text; routing delays stacked because clerks manually sorted by specialty keywords that missed nuance.
Our approach: NLP classification with strict de-identification, no generative free-text to patients, and evaluation reviewed by clinical admin leads. PHI never left the client's controlled environment.
What we built: ML model with confidence bands, human review queue for low-confidence items, audit logs and responsible AI documentation aligned to internal privacy policies.
What stayed human: Clinical triage, treatment decisions and patient communication remained entirely with licensed staff.
Illustrative outcome: Routing queue depth reduced by roughly one business day on average during pilot weeks — based on operational metrics, not a clinical outcome claim.
