Autonomous Agents Are Overhyped. Evaluation Is Underrated.
The "let agents run everything end to end" pitch keeps getting the attention. The boring evaluation and observability work is what actually decides whether a system survives production.
The Gap Between AI Spending and AI Value
Adoption is up and pilots keep getting scrapped at the same time. That's not proof AI doesn't work, it's proof most organizations are still funding the demo phase and calling it production.
Why Pixo's Content Pipeline Is a Graph, Not a Prompt
Why Pixo uses a graph pipeline, cached persona, dual-model routing, and swipe feedback — and what still needs building next.
Enforcing Multi-Tenant Data Isolation at the Database Layer
Inside HealthJourney: an agentic clinic platform where tenant isolation is enforced with Postgres Row-Level Security, not application WHERE clauses.
Why I Think AI Job Losses Are Being Underestimated
Having built the agentic systems that compress entry-level and support work, I don't think the optimistic "it's just task compression, headcount stays the same" story holds up.
Chaos Testing and AI-Generated Load Scenarios for APIs
Building a load-testing tool that turns an OpenAPI spec into test scenarios with AI, runs them through k6, and takes security seriously.
Two-Layer Semantic Matching for Career-Path Recommendations
How a Neo4j transition graph and a two-layer embedding match work together to recommend realistic career moves, not just similar job titles.
Why This Investor Recommender Uses Two GATv2 Models, Not One
Inside a GNN-based recommender that matches startups with investors: GraphSAGE candidate generation, dual GATv2 models, and why one shared model wasn't the right call.
Why Candidate Generation Comes Before Ranking in This Investor Graph
Why the Doriot investor recommender runs GraphSAGE retrieval before GATv2 ranking, then Bayesian re-ranking and rivalry signals.
When to Fall Back to a Reasoning Model Instead of Always Using the Bigger One
Inside the routing logic between GPT-4 and DeepSeek-R1 in a production RAG agent, and why the strongest model isn't the right default for every request.
Building a RAG Research Agent With Multi-Label Intent Routing
How I built a market research agent that classifies queries across 20+ intent categories before retrieval even starts, then routes between SQL and vector search.
Comparing SVM, ResNet, and EfficientNet for Player Contact Detection
An MSc computer vision project on detecting player contact in American football: why SVM wasn't the right tool, why augmentation fixed more than architecture did, and how YOLOX held up under cropping.
Designing Retrieval and Ranking for a Lyrics Search Engine
How a lyrics search engine combines exact-order pattern matching with TF-IDF, weights the two signals, and what we tried for fuzzy matching that didn't survive contact with 5 million tracks.
Security and Privacy Tradeoffs in a Card Rewards App
Building a card rewards app means handling financial data and real-time location at once. Why using Plaid instead of homegrown bank credential storage was the right call.