ML Systems Architect
With a decade of experience across enterprise-scale organizations and startups, I design production AI platforms spanning data pipelines, model training and evaluation, serving infrastructure, observability, and cloud architecture.
striff.io reviews the architecture of every pull request: it parses the repository into a typed dependency graph, scores the changed edges with a graph neural network, and posts deterministic structural findings as a GitHub check. This post breaks down the production pipeline behind it, and the three places we deliberately chose the boring option: a…
Read article ↗How striff.io uses a neurosymbolic pipeline with typed dependency graphs, Chidamber-Kemerer features, and an edge-prediction GNN to flag the dependencies in a pull request that carry architectural risk. Covers the graph construction pipeline, the 403-dimensional feature vector, why we switched from node-level anomaly scoring to edge-prediction, twelve deterministic detectors of which only five are allowed…
Read article ↗Lessons from building Fylo’s ingestion pipeline that turns scientific papers into typed discourse graphs. Covers ShEx schema as executable contract, a phased LLM extraction loop that closes the validator-LLM feedback gap, and a three-tier cross-document merge policy that keeps the graph converging as more papers are ingested.
Read article ↗After close to a decade working on ML infrastructure, including GPU clusters, autoscaling pipelines, and model serving systems, the transition into LLM-based production systems turned out to be less of a clean break than the hype suggests. The problems do not change so much as evolve, and they get harder in specific ways. This post…
Read article ↗Deploying a large language model is not the hard part. Deploying one that is safe to operate, cost-effective to scale, and straightforward to reason about under load is where most teams run into trouble. This post walks through an architecture developed at HADI Technology for running self-hosted LLM inference in production, using vLLM as the…
Read article ↗Teams that invest heavily in model development often ship the surrounding infrastructure as an afterthought — and pay for it later in operational failures that are difficult to diagnose and expensive to fix. This post distills a reference MLOps architecture built from repeated production engagements, using a GitHub pull request categorization pipeline with a PyTorch…
Shopify works until it doesn’t — vendor lock-in and transaction fees compound at scale, and the platform gives you limited control over infrastructure when it matters. This post walks through a CloudFormation-based reference architecture for running PrestaShop on AWS with the operational properties of a production system: auto-scaling application tier, managed RDS database, ElastiCache for…
Read article ↗REST and message queues are the two dominant approaches to inter-service communication in distributed systems, and teams frequently choose between them based on assumptions that do not hold under scrutiny. Synchronous HTTP calls are not always simpler or more reliable than async messaging, and message queues are not always the right choice when decoupling is…
Read article ↗Migrating a monolith to microservices is a difficult undertaking regardless of technical quality, but the difficulty scales dramatically with how coupled and procedural the source code is. When a codebase lacks clear object boundaries, the decomposition process becomes a guessing game about which pieces can be extracted without breaking everything else. This post demonstrates how…
Read article ↗Serverless apps depend on many managed services — storage, caches, load balancers, execution environments — which makes deployment automation non-trivial compared to a single application binary. Without structure, provisioning infrastructure and deploying code across dev, staging, and production environments becomes a manual, error-prone process. This post covers four practices for keeping that process automated and…
Read article ↗Code review comments contain signal about what engineers care about — naming, logic, performance, test coverage — but it’s buried in unstructured text. This post builds an SVM classifier to categorize 30,000 GitHub PR review comments by technical topic.
Read article ↗Swagger UI communicates endpoint details well but fails at conveying the shape of a complex API at a glance. When an API exposes dozens of resources and hundreds of operations, developers need a high-level map before they can navigate the detail. This post proposes a table-based documentation format that presents resources, operations, and their relationships…
Read article ↗Most OOP languages claim the label but miss what Alan Kay actually meant when he coined the term. Polymorphism, encapsulation, and inheritance are frequently cited as its pillars — but these exist in functional languages too. Kay’s actual vision was rooted in biology: autonomous objects communicating exclusively through message passing, with no direct access to…
Read article ↗A bash script that monitors local disk storage and reports usage to a Slack channel, color-coded by severity. Zero dependencies, configurable thresholds, deployable across multiple instances in minutes.
Read article ↗Relational databases were not designed for time series data — as write volumes grow, table cardinality climbs and query performance degrades in ways that are hard to tune around. Purpose-built time series databases like InfluxDB handle this workload efficiently by design, with compression, downsampling, and retention policies built in from the start. This post explains…
Read article ↗