📰 Dev.to · Daniel Romitelli

I built an AI message review console around a blunt rule: generation is allowed to suggest text, never to decide whether text is allowed to exist. The

I built a compliance audit sampling tool around a finite-population hypergeometric bound, then treated the confidence claim as something that could br

Part two of Validation Geometry. Seven live assets, every one miscalibrated, and the direction differs by asset: ETH over-confident at the tail, six o

A rewrite focused on the engineering pattern in Holographic, Law-Bound Memory: retrieval should return governed context with admissibility metadata an

How Neuroloq replaces static lesson completion with time-decayed mastery, stability scoring, and shared learner context across browser and phone tutor

A clinic messaging thread needs durable states, clear ownership, and channel policy before any chat interface can be trusted.

How I designed the video generation pipeline’s content-addressable lookup so retries reuse the same artifact while different model routes, seeds, and

I rewrote the embedding pipeline around a failure case I hit in production: a single blended embedding kept surfacing candidates whose experience l...

Before the generator writes a paragraph, a three-lane query fan-out, file-path-aware dedupe, a sufficiency threshold, and pinned excerpts decide wh...

I built a reward calibrator that sits above the scoring signals and searches for better weights instead of hand-tuning them by feel. It normalizes ...

How Neuroloq decides whether an attachment is document-like, which OCR mode to use, and why the router itself is the product.

I rewrote the post around the real startup path in `app/yapper.py`: the dependency check, the import order, and the shared settings path in `app/co...

I built a depth segmentation path that treats shape as the signal and texture as noise. The interesting part is the sequence: depth arrives, gets v...

A deep dive into `lib/scene-compiler/text-detector.ts`, showing how I normalize fal.ai and Florence-2 OCR output, classify subtitle and watermark r...

I built the handoff so the web app never has to guess what a segment, bounding box, or measurement means. The trick is to normalize geometry before...

I added a voice-activity control to Yapper tray mode because the app was still handing Whisper recordings that contained no speech. This rewrite st...

A production assistant that can’t resume state can’t own a workflow. After watching real teams lose hours to “AI amnesia” across resets, I built a ...

Part 1 of my “How to Architect an Enterprise AI System (And Why the Engineer Still Matters)” series: why I rebuilt my email intake path around dete...

In Part 2 of my “How to Architect an Enterprise AI System (And Why the Engineer Still Matters)” series, I explain why my production extractor runs ...

I learned the hard way that “just call one enrichment API” turns into 30–40% missing fields in production—and worse, it quietly teaches your system...

I designed the Outlook Add-in so the AI never writes straight into the CRM. It streams a draft into a human-reviewed form—field by field, with conf...

In Part 5 of “How to Architect an Enterprise AI System (And Why the Engineer Still Matters)”, I explain why I treat feature flags as a safety contr...

A worker fleet that retries safely is mostly about being explicit: atomic claiming with FOR UPDATE SKIP LOCKED, leases that can go stale and be rec...

I built a sync engine against a CRM API that behaved differently in production than its documentation implied. The result is a pipeline that auto-g...