Full-stack ownership from infrastructure to front-end
Solo product engineer redesigning a monolithic Python/C++ desktop application into a cross-platform, cloud-ready product — from architecture to deployment.
Replaced a tightly-coupled Python/C++ desktop app with a decoupled, cloud-ready product spanning six platforms.
Before — quidient-reality-python
- 1,799 files, 37,065 symbols, 333,522 call-graph edges
- wxPython + VTK + custom C++ raytracer in one process
- Windows desktop only — no client-server separation
- Duplicated app variants with copy-pasted, diverging code
- Hardcoded key bindings — code change required to remap
- No authentication, no cloud, no mobile, no updates
- Help system: a 2-line unimplemented stub
- Single-frame export only — no video rendering
- No measurement tools
- Manual PyInstaller builds, no CI/CD
After — Quidient Virtualize
- 238 files, 1,775 symbols, 7,397 edges — 87% fewer files
- React + Babylon.js + Tauri — composable, <700 LOC/file
- Windows, macOS, Linux, iOS, Android, Web
- Shared auth library eliminates all cross-app duplication
- Fully remappable controls with international layout support
- Azure AD OAuth2 + TPM auth, full AWS cloud stack
- 11-section help with live keyboard animations synced to video
- Keyframe video render tools: multi-output, 24–120fps, up to 4K
- 3D measurement with 6 unit types + scale calibration
- GitHub Actions CI/CD with Azure code signing
95% fewer symbols. 98% fewer call-graph edges. The new architecture achieves dramatically more functionality with dramatically less complexity — evidence of clean separation of concerns and elimination of legacy coupling.
Full Before/After Comparison Table
| Dimension | Before (quidient-reality-python) | After (Virtualize) |
|---|---|---|
| Codebase | 1,799 files, 37K symbols, 333K edges, duplicated variants | 238 files, 1.7K symbols, 7.4K edges, zero duplication |
| Frontend | wxPython + VTK, 1,300+ line monolithic panel | React + Babylon.js + Tauri, composable components |
| Coupling | Frontend embedded in backend process | Client-server WebSocket with binary streaming |
| Platforms | Windows only | Windows, macOS, Linux, iOS, Android, Web |
| Controls | Hardcoded in source | Remappable, persisted, international layout support |
| Selection | Voxel bounding box only | Freeform lasso → 3D neural primitive editing |
| Media Export | Single-frame PNG/TIFF/EXR | Keyframe video, multi-output, up to 4K@120fps |
| Help/Docs | 2-line stub (never implemented) | 11-section system with synced video + keyboard animations |
| Measurement | None | 3D tool with 6 unit types + scale calibration |
| Cloud | None | AWS CDK IaC, CloudFront, ALB sticky sessions |
| Auth | None | Azure AD OAuth2 PKCE + TPM device signing |
| Updates | Manual PyInstaller rebuild | Silent background via privileged Rust Windows service |
| CI/CD | Manual build + external exe deps | GitHub Actions, Azure signing, component change detection |
| Reconstruction | Blocking subprocess pipeline | WebSocket-controlled live training, interactive start/stop |
Four repositories, twelve applications — each designed, implemented, and shipped solo.
quidient-virtualize Flagship
app/
React + Babylon.js + TauriCross-platform 3D viewport client for real-time neural radiance field visualization. Ships to 6 platforms from one codebase.
- 9-DOF first-person camera + orbital camera with adaptive input detection
- Fully remappable controls with Keyboard Layout API for international keyboards
- Freeform lasso selection projecting into 3D neural primitive space
- Keyframe video export: multi-output, multi-res, 24–120fps
- 3D measurement tool with 6 units + scale calibration
- 11-section help system with live keyboard animations synced to video
- Depth-based frame reprojection for latency-hiding
python/
Python + asyncio + PyTorchWebSocket rendering backend hosting the proprietary Rust-based 3D engine via Python bindings.
- 100 concurrent connections with full middleware chain
- 25+ message type router across 6 handler categories
- Binary frame pipeline: Rust engine → JPEG/PNG encode → WebSocket stream
- Cloud/local mode switching with SQS + Lambda integration
- Nuitka standalone build with traceback encryption
launcher/
React + Tauri + RustDesktop scene manager and engine orchestrator — the entry point users install.
- Engine version management via named pipe IPC to update service
- Scene discovery with metadata display and launch orchestration
- Embedded OAuth2 redirect server (localhost:34567–35467)
- 1,500+ lines of Rust for engine verification, hash checking, IPC
update-service/
Rust · Windows ServicePrivileged Windows service performing elevated operations without UAC prompts.
- Named pipe IPC for updates, installs, copies, deletes, launches
- Secure pipeline: S3 download → SHA256 verify → extract → stage
- Desktop session awareness for user-visible process launching
cloud-cdk/
AWS CDKInfrastructure-as-Code for hosting the browser viewport and routing to GPU instances.
- Two-tier deploy: SPA-only (S3+CloudFront) or full (ALB+sticky sessions)
- ACM TLS, Route53 DNS, security group isolation
shared/ + CI/CD
TypeScript · GitHub ActionsCross-app auth library (@quidient/virtualize-auth) and full CI/CD pipeline.
- PKCE OAuth2, token refresh, JWT parsing, AuthGuard, TPM device auth
- Component-level change detection — only rebuild what changed
- Azure Trusted Signing for all executables and MSI installers
- Dev → production promotion pipeline with S3 + DynamoDB sync
qvd-cdk Backend API
Cloud Infrastructure
AWS CDK731-line CDK stack defining VPC, S3, DynamoDB, Lambda, API Gateway — the entire backend.
- VPC across 2 AZs with Gateway Endpoints for S3/DynamoDB
- Versioned encrypted S3 with lifecycle policies
- 3 DynamoDB tables with GSIs and TTL
- 13+ Lambda functions with least-privilege IAM
- REST API: 12+ routes, 100 req/s throttle
Delta Patching System
Cost OptimizationHash-based delta patching that transforms update distribution from linear-scaling to near-constant cost.
- SHA-256 per-file tracking across all build versions
- File-level diff: add/update/delete operations only
- Dual-hash strategy for code-signed builds
- Patch caching: DynamoDB TTL + S3 lifecycle
- 95–99% bandwidth reduction vs. full downloads
Auth + DevOps
SecurityTwo-layer authentication and automated build pipeline.
- Azure AD JWT authorizer via JWKS + TPM device signatures
- Canonical request signing with clock skew validation
- CI upload pipeline with dual-hash mapping + concurrent uploads
- Bug reports: presigned S3 upload → auto-create Asana task
Internal Tools 5 Apps
| Tool | Problem Solved | Impact |
|---|---|---|
| dent-measure-3d | Manual dent annotation from CloudCompare console logs | 20 hrs/car → ~2 hrs (10x) |
| dent-segmenter-3d | No panel-level dent classification workflow | Structured reporting + parallel annotation |
| synth-dents | No training data for dent detection ML models | Automated dataset generation (100s images/run) |
| quidient-keylogger | Manual timestamp entry for keyboard animation data | Made live help animations feasible |
| ios-data-upload | Manual iOS data transfer configuration | Single-click workflow |
dent-measure-3d was built in 2 hours and immediately saved ~18 hours per car in annotation time. Both dent tools were used from WP5 through WP8 on the Polaris contract.
quidient-reality Contributions
Config System
662 linesCreated TrainConfig + config_utils from scratch — 13 named presets so the product app can generate correct training configs without knowing research hyperparameters.
Build Fixes
Critical PathWorked with Brad Day to fix Cython runtime compile bug that broke packaged executables. Migrated track_builder from Cython to C++/pybind11.
Env Map + API
PipelineWired environment maps into the rendering pipeline. Refactored depth fusion CLI into callable API to avoid source exposure in packaged builds.
Concrete cost savings, time reductions, and capability gains delivered this year.
Bandwidth Reduction via Delta Patching
Hash-based delta updates deliver only changed files. At 500 users with weekly updates, this saves $22,000–$67,000/year in AWS transfer costs — and client downloads drop from minutes to seconds.
Dent Annotation Speed
dent-measure-3d and dent-segmenter-3d reduced per-car annotation time from ~20 hours to ~2 hours. Used across WP5–WP8 of the Polaris contract — Brandon Angelini (former Commercial BDEV Head) described the tools as "very integral to our deliveries."
Platform Reach
From Windows-only desktop to Windows, macOS, Linux, iOS, Android, and Web — from a single codebase.
Codebase Reduction
1,799 files → 238 files while delivering dramatically more functionality. 95% fewer symbols. 98% fewer call-graph edges.
Security & Auth
The legacy system had zero authentication. Now: Azure AD OAuth2 PKCE, TPM device signing, two-layer API auth, code-signed binaries.
CI/CD Pipeline
From manual PyInstaller builds to automated GitHub Actions with component-level change detection, Azure Trusted Signing, and dev-to-prod promotion.
DoD Contract Milestones Delivered
Delivered two Polaris contract milestones to the DoD: AVID (Q1 2025) and EBI Final Delivery (Sep 2025). Dent annotation tools were integral to contract deliveries across WP5–WP8.
Marketing Video Pipeline
Render tools built for the product are now the primary way the communications/marketing team produces LinkedIn and external outreach videos — described by the team as easier to use than Blender. Cross-team collaboration on surfel and luminel ERT visualization generated numerous promotional videos.
This wasn't a prototype. Each system is production-quality with edge cases handled.
Viewport Controls & Camera System
- First-person camera: WASD movement, Q/E vertical, Z/C roll — all 9 degrees of freedom
- Orbital camera: ArcRotateCamera for object-centric inspection
- Adaptive input detection via Strategy pattern — auto-switches between mouse+keyboard and touch
- Zustand-persisted keybinding store with schema migration across versions
- Keyboard Layout API integration for AZERTY, QWERTZ, and other international layouts
Selection Tool — 3D Neural Primitive Manipulation
- Freeform lasso selection in 2D that projects into 3D via camera intrinsics
- Backend raycasting: polygon rasterized into 2D mask, neural primitives projected into screen space
- Operations: Delete/Crop, Invert, Rotate (any axis, 1–180°), Color override, Save, Reset
- Accumulative selection + real-time re-render after every operation
Video Export — Keyframe Camera Track System
- Named camera tracks with keyframe capture and spline interpolation
- Multi-output: RGB, Albedo, Depth, Normals, Emission — separate video files
- 24/30/60/90/120 fps at 480p through 4K resolution
- Real-time progress dialog with phase tracking and per-frame counters
- Preview playback, loop option, auto-open output folder
Interactive Help with Live Keyboard Animations
- 580+ line video component with full physical keyboard layout as React components
- requestAnimationFrame-driven sub-frame-accurate sync between video and keyboard highlighting
- Adapts to user's custom keybindings — remapped keys highlighted correctly
- Mouse state visualization synced to video
- Keyboard Layout API for correct labels on non-QWERTY layouts
Delta Patching — Architecture Detail
- Every file tracked by path + SHA-256 hash in DynamoDB across all versions
- Dual-hash strategy: pre-sign hash for diff generation, post-sign hash for integrity validation
- Cached in DynamoDB (metadata + TTL) and S3 (7-day lifecycle) — computed once per version pair
- At 500 users with weekly updates: projected savings of $22,000–$67,000/year
- Client download time: 80–240 sec → 0.4–4 sec
WebSocket Backend — Middleware & Protocol
- Middleware chain: error handling → timing → logging → sanitization → validation → auth → metrics
- Message router dispatching 25+ types: rendering, training, scene, mesh, FOD, config
- Binary frame pipeline: Rust engine → OpenCV encode → JSON metadata + binary payload
- Cloud/local mode via env var — switches to SQS + Lambda session management
From legacy quidient-reality-python maintenance through complete product redesign, architecture, and alpha launch.
One engineer replaced a legacy monolith with a modern, cross-platform product — from first line of code to alpha launch in 11 months.
This body of work spans full-stack product development (React, Babylon.js, Rust, Python), cloud infrastructure (AWS CDK, Lambda, DynamoDB, S3), systems programming (Windows services, named pipe IPC, TPM authentication), CI/CD automation (GitHub Actions, code signing, delta patching), and machine learning tooling (synthetic data generation, COCO annotations) — all designed, built, tested, and shipped solo.