Memory deserves a place to rest.

AfterLight is a reflective, memory-grounded experience by Umbrella Research. Not a simulation. Not therapy. A quiet space for what matters.

<2sCold Start Latency
<35s3D Reconstruction
5,023Mesh Vertices
EU-OnlyData Residency

What is AfterLight?

AfterLight is a research-driven project exploring how memory can be organized, reflected on, and — optionally — represented through AI. It begins with conversation, not technology.

You share memories at your own pace. Ben, an AI reflection companion, helps you describe, organize, and find meaning in what you’ve shared. Everything is grounded in what you provide. Nothing is invented.

AfterLight is built by Umbrella Research — a small team focused on ethical, human-centered approaches to memory and AI.

Introducing AfterLight

A brief overview of the AfterLight project — what it is, how it works, and why it exists.


Platform Capabilities

Three ways to preserve what matters.

From subtle photo-based animation to full 3D reconstruction, each capability is designed to handle memory with care.

01

Micro Animation

Emotional Fidelity Engine

Neural micro-animation uses LivePortrait retarget MLPs to add subtle, natural movement to a still photograph. Gentle blinks, soft head motion, and lip nuance — all within ethical boundaries.

  • Micro-expression synthesis
  • Per-frame BiSeNet face compositing
  • Customizable parameters via conversation with Ben
  • 100+ tunable parameters with preset system
02

3D Echo

Spatial Reconstruction Layer

DECA face reconstruction with FLAME2023 topology produces a textured 3D mesh from one photograph. CodeFormer AI refinement enhances facial detail before geometry extraction.

  • Single-image 3D reconstruction
  • 5,023 vertices / 9,976 faces per mesh
  • 1024×1024 PBR texture with UV projection
  • Interactive Three.js GLB viewer
03

Research Avatar

Consent-Native Identity Layer

SMPL-X body fitting via PIXIE combined with 3D Gaussian Splatting produces anatomically accurate, high-fidelity 3D avatars from 8–20 photographs or a short rotation video.

  • PIXIE SMPL-X fitting (10,475-vertex body mesh)
  • 3D Gaussian Splatting (up to 20K iterations)
  • Adaptive loss scheduling with foreground masking
  • PSNR 24–25 dB on benchmark validation
Technology & Architecture

Built on research-grade infrastructure.

Every component is designed for privacy, performance, and ethical compliance — from neural inference to data residency.

Neural Pipeline

1
InputMaterial UploadPhotos and video uploaded to encrypted GCS storage with EXIF stripping and magic byte verification.
2
ProcessingNeural InferenceNVIDIA L4 GPU services run face detection, DECA reconstruction, LivePortrait animation, or Gaussian Splatting.
3
FittingSMPL-X Body ModelPIXIE encoder extracts body shape, pose, and expression parameters for anatomically accurate mesh generation.
4
TrainingGaussian SplattingAdaptive densification with confidence-weighted identity fusion and multi-scale reprojection loss optimization.
5
OutputSecure DeliveryGLB meshes, preview videos, and training logs delivered via time-limited signed URLs. No permanent public access.

Cloud Infrastructure

GPU ComputeNVIDIA L4Cloud Run GPU + Compute Engine g2-standard-8 in europe-west1. Scale-to-zero when idle.
Pipeline Speed< 35 seconds3D Echo end-to-end reconstruction. Research Avatar 90s–600s depending on iteration count.
Microservices4 Cloud RunBackend, website, neural animation, and 3D Echo — each auto-scaling independently in europe-west1.
Data ResidencyEU (europe-west1)All data processed and stored in Belgium. GDPR-compliant architecture with data export support.
EncryptionAES-256 + TLS 1.3Data encrypted at rest (Google-managed keys) and in transit. Security headers, CSP, and HSTS enforced.
Backup & RecoveryDaily automatedCloud SQL PostgreSQL 16 with daily backups, 14-day retention, PITR, RTO ≤30 min / RPO ≤5 min.
Ethics & Human-Centered AI

We don’t build what the person was. We build a space to sit with what you remember.

Every design decision is guided by six ethical principles. AfterLight is not a simulation of consciousness — it is a reflective tool, constrained by design.

Consent First

No avatar is ever created without explicit, cryptographically verified consent from the subject. Consent can be modified or revoked at any time.

Psychological Safety

Our deployment guidelines are co-developed with clinical psychologists and human identity researchers to prevent dissociation, misrepresentation, or harmful use.

Transparency by Design

Every rendered output is cryptographically watermarked. Our provenance API lets any observer verify whether content was AI-generated or human-captured.

Independent Review

An independent ethics board composed of AI safety researchers, bioethicists, and civil rights advocates reviews all high-stakes deployments before approval.

Investor Relations

Building at the edge of what’s possible.

Umbrella Research is completing its incubation phase and entering an accelerator program to scale AfterLight’s infrastructure, establish international operations, and bring memory-grounded AI to market.

1,250+Tests Passing
4GPU Services
3Neural Pipelines
EU-OnlyInfrastructure

Company Roadmap

Q1 2026Incubation CompleteCore platform built — 3 neural pipelines, 4 Cloud Run services, full EU infrastructure.
Q2 2026Accelerator EntryJoining accelerator program. Legal entity setup in Germany and Romania.
Q3 2026Office LaunchOpening offices in Germany and Romania. First research partnerships formalized.
Q1 2027Platform BetaPublic beta launch with full pipeline access for early adopters and research partners.
Q2 2027Commercial LaunchGeneral availability with enterprise tier, API access, and international expansion.
SEED ROUND · OPEN

Invest in the embodiment layer.

We’re raising seed funding to scale infrastructure, hire key research talent, and accelerate our path from incubation to commercial launch.

Get Involved

Join the research frontier.

We are selectively onboarding research partners, early adopters, and academic collaborators. Access to AfterLight’s full API stack is available to qualified applicants.

EARLY ACCESS PROGRAM

Get in touch

Tell us about your work and how you’d like to collaborate. We review every application.

We review applications weekly. Expect a response within 5 business days.