Nidus

Native iOS/SwiftUI iPhone application for mosquito-control (vector) field technicians. Technicians survey sites, record "nidus" (breeding-source) observations, inspections, and treatments — largely speech-driven: audio is transcribed, then parsed into structured knowledge graphs. Data syncs to a backend mirroring Placer's "FieldSeeker" system.

Built with SwiftData + SQLite (SQLite.swift + SQLiteMigrationManager), SwiftUI, MapKit/H3 hex cells, Apple's NaturalLanguage framework, and Sentry crash reporting (self-hosted GlitchTip).

Application entry

Nidus/NidusApp.swift boots a RootController/RootStore, configures Sentry, registers for remote notifications, and forces portrait orientation. Uses the classic controller/store pattern rather than pure SwiftUI MVVM.

Architecture

Code is organized under Nidus/:

• Controller/ — 11 controllers (Root, AudioRecording, Camera, Network, Notes, Region, Database, Settings, AudioPlayback, Error, Toast) coordinating stores/services.
• Store/ — SwiftData @Model persistence plus in-memory KnowledgeGraph (parsed field data), NotesStore, RegionStore, SettingsStore, audio recording/playback stores.
• Service/ — KnowledgeService (NL speech→knowledge extraction), NetworkService (actor: uploads audio/images, login, fetch-updates), DatabaseService (SQLite), plus camera/photo/movie capture, natural language, SPC observer, and device lookup services.
• Model/ — note types (NidusNote, MosquitoSource, ServiceRequest, TrapData, PictureNote, AudioNote, etc.) and FieldSeeker.swift (Codable entities matching the FieldSeeker backend schema).
• Database/ — SQLite schema, migrations, and a hand-written Query.swift layer.
• API/ — upload payloads (NidusNotePayload, ImagePayload, AudioPayload) and backend responses (IosClientResponse, FieldseekerResponse).
• View/ — SwiftUI views (map/breadcrumb/route/inspection-summary tabs, edit views, detail sheets, camera UI).

Data model / field capture

The domain is organized around the session tree: Site Survey → Safety/Access constraints → Observations → Features → Inspections/Treatments. Capture is dual-mode — voice triggers ("begin feature", "safety assessment", "negative observation") or tap navigation — and values are parsed from natural speech:

• Enum matching ("aggressive animal" → SafetyConstraint), pattern parsing ("twenty by forty by six" → dimensions), quantity+noun (groups), free-text into structured fields.
• Raw transcript is preserved as evidence but not shown to the technician during capture.
• A rule-based parser (KnowledgeService.ExtractKnowledge) — not ML — walks lemmatized tokens/grams to extract habitat, water conditions, dimensions/volume, density, life stage, genus/species, dip counts, treatment product/amount, and safety/access mitigators.

Backend sync

NetworkService handles login (server URL/credentials from settings), pulls incremental updates (IosClientResponse with a since timestamp returning service requests, mosquito sources, traps), and uploads audio files and pictures (with background URL-session support via BackgroundDownloadWrapper). Photos are attributed to the currently-active tree node.

Recent work / current state

The most recent artifact (2026-08-25) is a data-model & utterances package for a "Placer MVCD Phase 1" rollout: 20260825-placer-nidus-notes-data-model-package/ containing

• data-model.md — the canonical session-tree data model (entities, attributes, enums, triggers, recognition rules);
• utterances.yaml — 20 real technician utterances (10 structured + 10 conversational) with field extractions, covering dry features, access/safety constraints, and groups;
• utterances-review.md — field-by-field extraction breakdown per example.

This indicates the current focus is formalizing the data model and validating the rule-based extractor against real transcription data for the Placer Mosquito & Vector Control District Phase 1 rollout.

Notable observations (risks / debt)

• Two persistence stacks coexist: SwiftData @Model (NidusNote) and hand-rolled SQLite (DBSchema/Query.swift). Some types bridge both (e.g. NidusNote.toPayload() vs MosquitoSource Codable). Likely migration/debt hotspot.
• Typo/legacy residue: Store/KnowledgGraph.swift (misspelled filename); Note.swift defines a NoteCategory/Note protocol while Model/Note/ also has NoteProtocol.swift — two overlapping note abstractions.
• Hand-written extraction logic in KnowledgeService is brittle by nature (regex/gram heuristics, ordinal/cardinal parsing); the utterances corpus is meant to harden it.
• Tight coupling: RootController holds 11 sub-controllers and direct references to LocationDataManager; preview/test seams are thin (RootControllerPreview, forPreview helpers).
• Small unit-test surface (21 Testing-framework tests, mostly inspection*/mosquitoSource* extraction cases) relative to the parsing complexity.
• Stray non-source artifacts committed at repo root (.heif/.m4a media files, a .jpeg), plus two peripheral Xcode projects (SpeechTest, SpeechTrainer) alongside the main Nidus Notes.xcodeproj.

Bottom line

A mature, speech-first field-capture app with a sophisticated rule-based NL parser, currently formalizing its data model and validating extraction against a real technician utterance corpus for the Placer MVCD Phase 1 rollout. Main structural risks: dual persistence stacks, overlapping note abstractions, and coupling in the root controller.