Lazarus

Everyone has a place they would go back to. A kitchen in July. A street after rain. A voice calling you in for dinner. Lazarus finds that place in conversation, and builds it again.

Why it matters

56.8M

people need palliative care each year. About 14% receive it.

WHO

57M

people live with dementia, with nearly 10 million new cases a year.

WHO · 2021 data

~30,000

Israelis in rehabilitation since 7 October 2023; 63% with psychological injury.

Israel Ministry of Defense · 2026

The Memory Engine

One conversation in. One world out.

Four layers of AI do the work of a biographer, an archivist, a set designer and a session guide. A clinician stays in charge of every scene.

01Listen

A biographer that never runs out of time.

A voice AI interviews the patient and family the way a skilled biographer would. It hears fatigue, never pushes, and stops on a single word.

Speech + language models · Hebrew, Russian, Arabic, English

02Understand

Stories become structure.

Every answer joins a private memory graph: place, season, people, light, sound, smell, feeling, and the topics that must never be raised.

Clinical memory schema · consent ledger · family-owned

03Rebuild

A place, rebuilt from memory.

Family photos become 3D spaces. Generative world models fill in what the camera never saw. A fidelity dial runs from photoreal to dreamlike.

Gaussian splatting · world models · neural rendering

04Guide

A session that follows the person.

In real time the engine paces light, sound and narration to voice, breathing and pulse, closes gently, and records how the person felt.

Voice · breath · pulse · outcome measures

Guardrails in every layer

A clinician approves each scene
Red-listed topics are never opened
Risk language reaches the care team at once
Designed to run on site

Places people go back to

Every memory has an address.

From one conversation the engine rebuilds the place itself, then lets a person stay there for a while. Four illustrative memories, generated live in your browser. Turn the dial to see how sharp a memory should be.

Rebuilt from memory · illustrative

Grandmother’s house · July

“Summers at my grandmother’s. Apricot jam on the stove, and the window open to the yard.”

Placeher houseSmellapricot jamFeelingsafe

Fidelity dial

DreamlikePhotoreal

The Biographer, listening

It asks for the details that bring a place back.

LazarusYou mentioned summers at your grandmother’s. What did her kitchen smell like in the morning?

PatientApricots. She made jam, and the window was open to the yard.

Placeher kitchenTimeJuly, morningSmellapricot jamLightopen window

LazarusWhat could you hear from that window?

PatientHer radio. The neighbours’ kids. I felt so safe there.

Soundradio, childrenFeelingsafeAvoidset by the clinician

Illustrative example, not a patient transcript.

One engine. Three skies.

The same engine adapts to three very different rooms, each with its own clinical team and its own pace.

Hospices · palliative wards

A last good day.

Life review, a return to a loved place, a way to say goodbye. Built with the palliative team and the family, at the bedside.

Care homes · geriatric centres

Recognition, not recall.

Reminiscence for people living with dementia. The family leads; sessions are short, calm and seated, with a carer present.

Trauma clinics

Somewhere safe to stand.

Positive memories as a stable base for people with severe PTSD, added to therapy and supervised by the treating clinician.

Safety by design

Under supervision. Always.

An AI biographer, not an AI therapist. These rules live inside the engine, so they hold in every session.

Consent first

Patient or guardian consent, and the person’s own assent at the start of every session.

Clinician in the loop

Every scene is approved before anyone sees it, and the care team can stop any session.

Red lists

Topics the care team rules out are never opened, in conversation or in a scene.

Escalate, never treat

Risk language goes to the care team immediately. The AI does not diagnose.

Memories stay close

Designed to process on a bedside unit. The memory graph belongs to the family.

No synthetic departed

Relatives who have died appear only if the family and the clinician agree.

Where it runs

Any screen that can hold a memory.

VR headsetsCommercial headsets for the first pilots.

TabletFamily interviews and the scene preview a clinician approves.

TV or projectorGroup reminiscence in care homes, no headset needed.

LUMEN visorOur lightweight visor for patients lying in bed. Concept. Turn it in 3D ↓

LUMEN L1 · concept hardware

Made for a head on a pillow.

A light, tethered visor for patients who cannot sit up. Turn it, lift it, take it apart. The model runs live in your browser from the same geometry as our prototype files.

drag to rotate

Take apart 0%

Lift to return 0°

Display and optics

Panels: 2 × 1.3-type micro-OLED, Sony ECX344A
Resolution: 3552 × 3840 per eye, up to 90 fps
Panel range: Up to 96% DCI-P3, 6.3 µm pixel pitch
Lenses: Pancake modules, Ø40 mm
IPD: 58–72 mm, motorised
Eye tracking: 2 IR cameras and an IR LED ring per eye

Sensing and sound

Passthrough: 2 RGB cameras on the mainboard
Motion: IMU for head tracking
Microphones: 4 on the visor, 6 far-field in the Base
Sound: Open-ear speakers in the temple pads
Voice line: Light strip behind the glass that moves while the AI speaks
Exit: Lift the visor 0–75°; a Hall sensor fades the scene back to the room

Comfort in bed

On the head: ≈ 190 g design budget, no battery
Straps: 3D-knit crown loop; the rear loop unclips for bed mode
Bed mode: Nothing between the back of the head and the pillow
Light seal: Shore 30A silicone on magnets, one per patient
Fit: Ratchet fit dial on the right side
Size: Visor 188 × 60 mm; 200 × 170 × 248 mm with loops

Bedside Base

Compute: RTX 50-class GPU, CPU and NPU
Job: Renders the memory and runs the AI interviewer at the bedside
Link: USB-C tether with DisplayPort Alt Mode
Audio: 6-mic far-field array and a speaker
Dock: Charges and stores the visor; voice light ring
Size: Ø148 × 286 mm

Bill of materials · 12 parts · ≈ 192 g on the head

No	Part	Material and process	Qty	g
01	Front glass, black mirror	Smoked PC 1.4 mm, thermoformed, AR and anti-fog	1	12
02	Mainboard and passthrough cameras	Curved flex-rigid PCB, 2 RGB cameras, IMU, 4 mics	1	22
03	Visor shell and front halo	PA12 MJF for the prototype, then injection-moulded PC/ABS, matte white soft-touch	1	38
04	Micro-OLED panels	Sony ECX344A, 1.3-type, 3552 × 3840	2	10
05	Optics bezel and eye tracking	Black PA12; IR eye camera and IR LED ring per eye	1	9
06	Pancake lens modules	Off-the-shelf modules, motorised IPD 58–72 mm	2	36
07	Light seal	Shore 30A silicone, magnetic, swappable per patient	1	14
08	Crown loop	3D-knit spacer fabric, padded, clips to the side arms	1	14
09	Rear loop	3D knit, detaches for bed mode, tensioned by the fit dial	1	12
10	Fit dial	PC/ABS ratchet spool, right side	1	4
11	Temple pads and open-ear speakers	Silicone foam over 2 open-ear drivers	2	9
12	Tether, magnets, lift sensor	USB-C / DP Alt Mode tether to the Base; Hall sensor	1	12
	Head-mounted total, design budget			≈ 192

Finish. Matte white soft-touch shell, smoked glass with a black-mirror coat outside, a cool-white voice line behind it, light-grey knit loops with white clips, warm-grey seal. The Base is matte white with a smoked dome over the microphones.

Concept EVT-0. Weight, sizes and the GPU class are design targets, not measured values. Panel figures are Sony's published specs for the ECX344A.

Seed round · 2026

Help us bring Lazarus to its first hospice.

We are raising a $3M seed for clinical studies, the Memory Engine and the first pilots in Israel. Lazarus is in development.

Co-Founders: Michael Mazhara and Maiia Krylova · Netanya, Israel

mikemazhara@gmail.com