A single, honest timeline of every shipped release, every meaningful engineering decision, and every compliance change. We do not have a marketing blog — this is the work, raw and unembellished.
Resolves a long-standing memory regression when importing large GPX files. Topographic map rendering now uses a tiled vector pipeline that streams contour data per viewport — typical load dropped from 4.1 s to 680 ms on 3-year-old devices. Offline POI database rebuilt with 2026 OSM snapshot (12.4 M entries compressed to 187 MB).
Our on-device OCR pipeline (used in ReceiptVault) runs a 7-layer custom CNN at 11.4 ms median on an A14 Bionic — within a frame budget for 60 Hz preview scanning. This note walks through the quantisation strategy, the Core ML conversion pipeline, the latency budget allocation, and the profiling methodology with Instruments + our custom trace system. We share the actual MLmodel spec and a Swift snippet that runs the loop synchronously on a dedicated queue.
Per Article 24 of the Digital Services Act, we publish our quarterly transparency report covering content moderation actions, notice-and-action turnaround, complaint handling, and automated tooling accuracy. In Q2 2026: 41,820 user reports received across hosted platforms, 92.7% actioned within 24 hours, 0.4% escalated to out-of-court dispute settlement. Full machine-readable dataset published to our compliance portal.
Updated the market value estimation model for books, vinyl, and watches with Q2 2026 sales data from seven primary marketplaces (eBay, Discogs, Chrono24, AbeBooks, Biblio, StockX, GOAT). Median accuracy improvement: +6.3 percentage points across categories. Local model — no telemetry, no server-side training, no third-party pricing API call.
Yes, we tried Swift's new Codable synthesis improvements in Swift 6.1. Yes, we measured. No, we did not migrate. This note explains why our handwritten decoders for the Cadence behavioural-rhythm data model are still 38% faster on cold paths and produce 61% smaller binaries after dead-code elimination. Sometimes the obvious modern tool is not the right tool.
We welcome Jun Tanaka as Lead iOS Engineer (9 years of native iOS and watchOS, three Apple Design Award nominations) and Sofia Costa as Privacy & Compliance Lead (former EU data-protection authority counsel, CIPP/E and CIPP/US certified). The studio is now 14 people. The next planned addition is in Q1 2027.
Our annual ISO/IEC 27001 surveillance audit was completed on 14 May 2026 by an accredited certification body. No non-conformities raised. Three observations logged and addressed within 10 business days. The full Statement of Applicability (v3.2) is available to enterprise clients and partners under NDA.
A long-anticipated release. The writing vault now clusters fragments by semantic similarity using an on-device sentence-transformer model (22 MB, distilled, INT8-quantised) and surfaces a mood-based theming system that adapts typography and accent colour to the emotional register of the entry. Full feature article in the Inkdrop 4.1 release notes.
We have moved all client ad mediation configurations from legacy waterfalls to in-app bidding across AppLovin MAX, Google AdMob, Unity LevelPlay, and Meta Audience Network. Early client results: eCPM uplift of 18-34% on rewarded video, 12-21% on interstitials, with no statistically significant change in fill rate. Technical breakdown of the waterfall-to-bidding migration in the article.
Effective 22 April 2026, all consumer apps we ship default the CPRA "Limit the Use of My Sensitive Personal Information" and VCDPA "Opt out of sale or sharing" toggles to ON. Users in those regions now have to explicitly opt in to data monetisation rather than opt out. Region detection uses the device locale combined with the user's manually-set country override.
The HRV-based readiness engine now supports wrist temperature (Apple Watch Series 8 and later) and respiratory rate as inputs, improving recovery prediction accuracy by 9.2% in our 12-week validation cohort. The new periodised cycle planner lets coaches and self-directed athletes programme 6- to 16-week blocks with deload weeks auto-calculated from TRIMP load.
On 4 April 2017 we shipped our first app. Nine years later, we are 14 people, 14 live products, and we still care about the same things we cared about on day one: minimalist design, privacy by architecture, and experience above all. Thank you to every user, every client, every team member past and present.
After 18 months of internal use, we are releasing Confetti as an open-source library under Apache 2.0. Confetti replaces surveillance-style analytics with locally-aggregated counters that are k-anonymised and differentially-private before any optional sync. The result: 95% of the product insight, zero of the surveillance. Repo, docs, and Swift Package available today.
The Digital Personal Data Protection Act, 2023 became enforceable on 18 March 2026. We updated the region-adaptive compliance engine to detect Indian users via locale, apply the consent-first model, localise data residency for sensitive categories, and present the DPDP-aligned privacy notice. Children's data (under 18) is blocked by default for any data collection outside the strictly-functional baseline.
The new export pipeline produces AES-256-encrypted PDFs at an average of 108 KB per page (compared to 410 KB previously) while maintaining OCR text-searchability. The compression uses a custom JPEG2000 variant tuned for document content, and the encryption key is wrapped to the device Secure Enclave / Android Keystore.
We will be presenting at WWDC26 (San Francisco, June), Android Makers (Paris, April), DSA Workshops (Brussels, May), and the Audio Developer Conference (London, November). Elena will keynote at the Privacy Engineering Conference in Berlin on 18 September. Reach out via the contact form if you would like to schedule time during any of these events.
One email per month. Plain text. No tracking pixels, no marketing automation, no "engagement metrics." The same changelog you see above, plus a few thoughts on what we are reading and building. Unsubscribe with one click — we do not save your address anywhere except the list.
