Simor Consulting
Category: Operations
AI projects have a scoping problem. Teams either scope too loosely, "use AI to improve customer experience", or too tightly: "build a transformer model with 12 attention layers for intent classificati
Data pipelines break in production more often than they should, and the breakage is expensive. A pipeline that silently produces wrong data for three days before anyone notices has corrupted downstrea
Post-mortems are useful. Pre-mortems are cheaper. A post-mortem tells you why a project failed after the money is gone. A pre-mortem tells you why a project might fail while you can still change cours
LLM costs are easy to start and hard to control. A team ships a feature that calls GPT-4, the feature works, users like it, and the invoice climbs 15 percent month over month. The cost is not a proble
Most data mesh initiatives fail not because the architecture is wrong, but because nobody can answer the question: who owns this data product? When ownership is ambiguous, quality drops, SLAs go unmet
Teams approve LLM projects based on per-query cost estimates, then get blindsided by the actual invoice. The gap between estimate and reality is not a rounding error. It is a structural problem: the e
The decision to centralise AI infrastructure into a platform team usually comes after a period of decentralised pain. Three product teams independently built model serving pipelines. None of them shar
Most AI systems in production operate with observability that was designed for traditional software. Teams monitor CPU, memory, network, and error rates. These metrics tell you whether the server is r
An architecture review for an AI system catches design flaws at the cheapest possible stage: before implementation. A data pipeline that cannot handle the expected volume, a model serving architecture
Data pipelines break because data producers and data consumers have different assumptions. The producer assumes the consumer can handle null values in a column. The consumer assumes the column is neve
A model registry is the version control system for your trained models. Without one, teams track model versions by filename, store artifacts in ad-hoc cloud storage locations, and discover which model
AI vendor procurement is where organisations make binding commitments that are expensive to unwind. A three-year contract with a model provider locks you into their pricing, their rate limits, their m
Vector database capacity planning fails in predictable ways. Teams estimate storage based on vector count alone and discover at 60% capacity that memory consumption is growing faster than disk because
AI systems fail differently than traditional software. A traditional software bug produces incorrect output deterministically. The same input always produces the same wrong output, and a fix eliminate
Professional services firms sell judgment, billed by the hour or by the matter. That makes them both the biggest winners and the most cautious adopters of AI. The upside is real: every firm carries ho
The pipeline runs every night at 2 a.m. Nobody fully understands it. The original author left in 2019. It is part SAS, part shell, part stored procedures, and part a spreadsheet someone emails in. It
A head of ML at a 120-person company told us recently that his team had spent nine months trying to stand up a "proper MLOps platform." They had evaluated three orchestration tools, designed a feature
Most RAG systems are evaluated with vibes. An engineer runs ten queries, eyeballs the results, and declares the system "working." Three months later, a customer reports that the system confidently ret
The first Head of AI Engineering at a company inherits one of three situations. Situation one: there is no AI team, no AI infrastructure, and the mandate is to build from scratch. Situation two: there
Software rollbacks are well-understood. You deploy a new version, detect an issue, and roll back to the previous version. The rollback is atomic: the entire application reverts to the previous state.
On a Tuesday at 2:14 PM, a major model provider began returning elevated error rates for a specific model endpoint. By 2:31 PM, a customer support platform that depended on that endpoint was producing
Organisations that skip readiness assessment before investing in AI tend to discover their gaps expensively. A financial services firm spent four months building a customer churn prediction model only
Bias in AI systems is not a theoretical risk. It is a measurable property that can be detected, quantified, and mitigated at every stage of the pipeline. The teams that treat bias as an audit problem
The case for streaming is straightforward: data that arrives in minutes instead of hours enables decisions that were previously impossible. Fraud detection catches transactions before they clear. Pers
LLM inference costs follow a pattern that catches teams off guard. The first prototype costs almost nothing: a few hundred dollars a month during development. The pilot scales to a few thousand. Produ
Most data quality initiatives fail not because teams lack tools, but because they measure the wrong things. Teams track hundreds of data quality metrics, generate dashboards full of green indicators,
Prompt management in most AI teams starts the same way. One engineer writes a prompt, it works well enough, and the prompt gets committed to a config file. Three months later, there are forty prompts
Every AI infrastructure team eventually faces the same argument. One faction wants to build a custom solution because the commercial options do not handle their specific requirements. The other factio
Most vector database selection failures come down to one mistake: picking the technology before mapping the workload. Teams benchmark embedding search speed on a curated dataset, pick the fastest opti