DevSight

The problem

Investigating a new deviation meant manually keyword-searching four years of Trackwise records — roughly 30 minutes per question, and brittle: the right precedent was easy to miss if it used different wording. Investigators needed semantic recall over 2,213+ resolved cases without leaving the 30-day closure window.

Architecture

Reliability and scale

• The final v2.7.0 release removed roughly 2,200 net lines while adding dark mode, an accessibility pass and a tokenized component library
• The vector index is shared: BioTrack reuses it for a search scoped to bioburden investigations

What it did

DevSight (formerly TrackLA) was a hybrid-RAG assistant over four years of Trackwise deviations, 2022–2026: 2,213+ resolved cases chunked into 31,577 searchable units. It replaced ~30-minute manual keyword searches with ~2-minute semantic ones — a 93% time reduction — and surfaced the actual precedent cases behind every answer, so an investigator could check the source instead of trusting the summary.

Two-stage hybrid RAG

Search ran in two stages: a date-metadata filter narrowed the candidate set, then semantic vector search ranked across 10 Trackwise fields. Date filtering happened post-search, client-side — a deliberate workaround for Databricks Vector Search not handling DATE range operators cleanly.

• Conversation memory across the last 10 turns
• Real-time token streaming with voice input and read-aloud
• Document upload limited to Word and Excel — image and PDF disabled on purpose, to preserve streaming
• Mobile-optimized with a fixed input bar, for use away from a desk

The final release

v2.7.0 was a design-system rewrite that removed roughly 2,200 net lines while adding dark mode, an accessibility pass and a tokenized component library — the codebase got smaller as capability grew. It also introduced a single APP_VERSION constant as the source of truth for every version string, and self-hosted fonts because the plant floor blocks CDNs.

Why it was decommissioned

DevSight was decommissioned in 2026 during a global application consolidation: the site moved onto standardized global applications, and locally built tools in that space were wound down regardless of how they were performing. That was a portfolio decision, not a verdict on this application. At the time it held a self-reported 4.8/5 satisfaction score, was still delivering the search-time reduction it launched with, and had shipped its largest release — v2.7.0, the final one — in its last month of operation.

What survived

The capability outlived the app. On 2026-08-14 the general-purpose search was merged into the Horizon v1.2.0 proof of concept as pure R utilities plus one Shiny module with its own tests, and BioTrack uses the same vector index for a search scoped to bioburden investigations. Retiring an application and retiring a capability are different decisions, and only the first one was made.