The engineering stack behind our products
We build each capability once and deploy it across every product. The domain changes — fashion, QA, education, health — but the engineering primitives stay the same.
Four capabilities that power every product
These are the engineering primitives that appear in every system we build. The combination and tuning varies by domain; the underlying capability does not.
LLM Orchestration + RAG
Large language model pipelines wired to domain-specific retrieval stores. Responses draw on real operational data — product catalogues, test artifacts, clinical records — instead of relying on generic training knowledge.
In Selora, this powers language-driven QA analysis, scenario authoring, and reporting tied to real run data.
Computer Vision + Recommendation
Multi-stage visual pipelines that classify, tag, compare, and match images. Paired with recommendation models that learn user preferences over time and surface the right output contextually.
In Dress In Style, this combination powers wardrobe auto-tagging, visual understanding, and outfit recommendations.
AI Automation + Self-Heal
Autonomous execution layers that run high-volume repetitive work. When selectors drift, schemas change, or APIs fail, the self-heal subsystem detects the break and patches the pipeline without waiting for a human.
In Selora, this shows up as flaky-test control, automated triage, and self-healing when brittle selectors drift.
Human Control + Observability
Review loops, approval gates, audit trails, and real-time dashboards. Every autonomous action is logged, every critical path has a human checkpoint, and teams can observe system behaviour in production.
QA dashboards in Selora and review checkpoints in future clinical pilots keep humans accountable for the final decision.
How the stack runs inside each live product
The shared capabilities are configured differently per sector. Here is how the engineering comes together inside each live product.
Every item added to the wardrobe passes through a CV classification layer that extracts color, category, season metadata, and pairing signals — making the entire collection instantly searchable and matchable.
The style identity system learns aesthetic preferences from onboarding inputs and behavioral signals, continuously refining outfit recommendations to match how each user’s taste evolves.
A body-photo-based virtual try-on system overlays selected outfits onto the user’s image, enabling purchase decisions and outfit curation without physical try-on.
Selora turns recordings, existing automation, API flows, and AI-guided exploration into a framework-agnostic QA layer that can generate and run coverage across multiple frameworks without rebuilding everything from scratch.
Every run arrives with logs, screenshots, videos, traces, and failure context. When selector drift or brittle assertions break coverage, the self-heal layer can repair the test and push QA back toward green faster.
Visual review, flaky detection, readiness dashboards, and integrations with GitHub, Jira, Slack, and CI keep quality work connected to real ship decisions instead of scattered across disconnected tools.
Why this architecture works
A wardrobe categorization pipeline and a test canonicalization engine look like different systems on the surface. Under the hood, they share retrieval, vision, automation, and human-control primitives. That structural reuse is what lets us enter new sectors — education, health — without rebuilding from scratch. Each deployment inherits proven capabilities and adds domain-specific configuration on top.