Engineering lab / R&D

Summary

Infomerx builds in-house, browser-native technical R&D where the hard problem is real-time UX under load: heavy compute, tight feedback loops, and interfaces that stay responsive while graphs, audio, or 3D geometry evolve. Two active threads are collected on this page—both pre-release, described at architecture and intent only, with no stable public URLs linked from the marketing site yet.

Trama Audio Graph targets forensic audio restoration. A node graph keeps DSP stages explicit so operators can show demonstrable reconstruction step by step—not a single opaque pass.

Parametric wing playground explores flying-wing geometry, section-level aerodynamics, and mission-style scoring toward optimization in a SvelteKit + Three.js playground, backed by a local FastAPI service (including NeuralFoil-class airfoil analysis). NeuralFoil’s fast surrogate section models are used so aero can come back quickly enough for real-time, design-time feedback while stations, twist, and airfoils change—without blocking the editor on full CFD loops. The roadmap for wing optimization is to lean on that near–real-time NeuralFoil loop while editing, then introduce a two-phase optimizer where later stages call progressively higher-fidelity CFD only on a filtered shortlist of candidates—using aggressive filtering and selection up front so expensive simulations run where they earn their wall-clock. Prandtl-style bell spanload is an optional mode that differentiates the tool from generic wing builders; other spanload prescriptions (elliptic, uniform, custom, or hybrid targets) remain in scope on the same parametric core. Full-wing integration and optimizer endpoints are still in active development—this page does not assert flight-validated or certification-ready performance.

Why this matters to clients

These threads are not shrink-wrapped products you license today—they are proof of technical range in the same problem classes client teams care about: heavy browser UX, tight feedback loops between analysis and UI, and disciplined separation of concerns so systems stay explainable as they grow. If your procurement question is “can this practice handle demanding interfaces and services without losing the plot,” this page is intentionally specific so you can judge depth without marketing fluff.

Challenge

Forensic restoration work needs both heavy spectral tooling and a way to justify what changed: which stages ran, in what order, and how each contributed to the output. Browser audio hits latency, memory, and GPU vs CPU trade-offs at the same time—while the UI must stay fluid for live waveforms, zoomable spectrograms, and graph editing without fighting the audio thread, and without collapsing everything into one monolithic processor that cannot be inspected or extended.

Approach

Separate UI, CPU, and GPU concerns explicitly: Svelte owns interaction and graph editing; workers and WASM own predictable DSP paths; WebGPU owns throughput-sensitive spectral paths where supported. Favour deterministic state and clear boundaries so debugging stays tractable as the graph grows.

Build focus (Trama)

• Forensic restoration focus — Trama targets damaged, noisy, or compromised recordings where iterative cleanup and spectral repair are part of the workflow.
• Node graph as audit trail — Filters, envelopes, mixers, and spectral tools connect in a visual graph with real-time preview so reconstruction stays traceable and demonstrable.
• Dual compute path — WebGPU-first spectral pipeline with worker + WASM fallback for environments where GPU compute is limited or unavailable.
• Spectral editing affordances — Spectrogram viewing, selection, and masking flows aligned with restoration—not decorative viz.
• Developer ergonomics — Typed workers, WASM build integration, and structure that supports iterative DSP experiments without fork-lifting the whole app.

Parametric wing playground (in development)

Designing and iterating parametric wings in the browser means keeping 3D editing, analysis callbacks, and future optimization loops from tripping over each other: the UI must stay interactive while services return structured results the UI can trust and display. NeuralFoil-backed section analysis is chosen in part because it is fast enough for design-time loops—operators get near–real-time reactions as geometry changes—without pretending unfinished wing-level or optimizer paths are production-grade. The intended optimizer path is two-phase: wide exploration on fast surrogates, then progressively more accurate CFD on a filtered set so total simulation time stays under control.

• Interactive geometry — Three.js–backed editing and visualization tied to a single parametric wing model (planform, span axis, twist, airfoil stations).
• Local analysis service — FastAPI service for health checks, batch-style evaluation envelopes, and NeuralFoil-backed airfoil section analysis tuned for real-time design-time feedback (fast surrogate evaluation while the wing is edited); wing-level and optimizer candidate paths are still being wired toward the same contracts.
• Spanload flexibility — Optional Prandtl-style bell spanload steering as a product differentiator; other spanload targets and scoring terms stay on the same core—no forked “bell-only” architecture.
• Optimizer roadmap — Near–real-time NeuralFoil-backed feedback during design; later a two-phase workflow that reserves progressively higher-fidelity CFD for candidates that survive filtering and shortlisting, minimizing time in the slowest solvers.
• Roadmap realism — Mission scoring, ranking, optional browser flight validation, and high-fidelity tiers stay behind documented pipeline stages; marketing copy stays clear of airworthiness or certification claims.

Outcomes

Success is measured in engineering leverage for both threads: surfaces where new nodes, rendering strategies, or parametric experiments can be tried without rewriting the shell, with performance and provenance inspectable through the stack. For Trama, that means reconstruction you can walk through—graph in, demonstrable stages out. For the wing playground, it means a coherent path from editable geometry to scored candidates—surrogate feedback first, then a disciplined funnel toward higher-fidelity checks—without overstating what each tier proves today. Public demos and links for either product will be added here when they are stable enough to share.

Public demo

Not published yet. Neither Trama Audio Graph nor the parametric wing playground has a stable public build or hosted URL from this case study today; both remain in-house R&D until a deliberate release.