ABOUTFOSTY PRODUCT DESIGN ENGINEERING OUTCOMECRUX VISION MOVEMENT REVIEW VISUAL OVERLAY ENGINEERING OUTLOOKVAL CONTRIBUTIONS PRODUCTIONINHERITANCE CHALLENGE ENGINEERING IMPACT

EVAN LUEBBERTFOSTY

ABOUT

I build software I believe in.

NYC · Open to full-time and freelance work.

APPROACH

Software engineering and design is my superpower. It enables me to wield technology to build a better world.

Being able to solve my own problems is a luxury.

The ability to craft solutions for others is a privilege.

I’m devoted to making well-designed tools to tackle complicated problems. I embrace curiosity, explore creative solutions, and fill the gaps where software can make a difference. Everyone has used frustrating and poorly designed software. I can fix that. I build clear, seamless workflows around the way people actually work, so software can be an asset instead of a liability.

I LOVE BUILDING

Full-stack applications for complicated workflows
Intuitive and satisfying interfaces delivering polished UX
Innovative technology for projects with a positive impact
Data-heavy tools and visualizations
Software where reliability and trust matter

FOSTY

FOUNDER · FULL-STACK ENGINEER · PRODUCT DESIGNER

Fosty

2025 TO PRESENT

Built for animals in need and the people devoted to helping them.

Fosty is a foster coordination platform I founded and built for animal shelters and rescues. Before moving to NYC, I fostered 34 kittens through Colorado Kitty Coalition. I watched their team struggle with urgent care coordination split across texts, emails, DMs, and messy Google Sheets.

That friction can have a serious impact on outcomes. An animal in a critical condition deserves timely intervention, and disorganized communication can have tragic consequences. I saw a broken system and I knew I could build them something better. I was inspired to create Fosty, a custom platform to organize foster communication and record keeping. I reached out to the rescue with the proposal for Fosty, and they were ecstatic. I will always remember their response:

“We need you!”

FOSTYTHE PRODUCT

A dedicated platform for foster care

Animal rescues coordinate care across staff and volunteer foster homes. Fosty brings intake, records, foster requests, assignments, and real-time communication into one place, so teams can spend less time piecing together information and more time caring for animals.

FOSTERS NEEDED
The Fosters Needed page gives volunteers one place to see animals waiting for a foster home. Essential details are visible at a glance, helping volunteers quickly understand where they can help.
ANIMAL DETAILS
Each animal profile brings care needs and history into one clear record. Volunteers can understand the level of care required before offering to foster, helping rescues make better matches.
GROUP DETAILS
Animals that should stay together, such as a litter of kittens, can be managed as a group. Shared care information stays consistent while each animal keeps an individual record.
INTAKE
Intake is designed for the unpredictable flow of rescue work. Staff can enter animals in batches and create useful records even when details are incomplete, so care is not delayed by missing information.
REQUEST
Volunteers can request an animal or group directly from its profile. This turns interest into a clear, trackable step instead of another message staff must piece together.
ASSIGNMENT
Staff can review every request in one place and match animals with the right foster homes. Decisions stay visible, reducing confusion and helping animals move into care faster.
COMMUNICATION
Each foster home has a dedicated group chat with rescue staff, so questions can reach the right person without leaving anyone out of the loop. Animal and group tags, requests and assignments updates, and shared photos keep each conversation connected to the care it supports.

FOSTYDESIGN

Designed with the rescue team

A Fosty usability review call showing a structured workflow checklist alongside participants in Google Meet. — USABILITY REVIEW
A live usability review focused on friction, discoverability, and efficiency across core foster workflows.

The goal of Fosty is to provide features and workflows that fit naturally into existing rescue operations, reducing administrative chaos and helping teams make faster, clearer care decisions. There was no existing animal foster-care platform to use as a blueprint, so I had to invent the product model, workflows, and interaction patterns for this problem space, translating CKC’s fragmented real-world process into a structured, coherent system.

The platform serves two distinct user groups: coordinators who need operational visibility across many volunteers, and foster caregivers who need to see their assignments and stay in contact with staff. I worked directly with the rescue team to identify pain points, define requirements, and scope solutions that could be built and tested.

Fosty’s data model is designed around the core relationships in rescue operations. Animals can belong to groups such as litters, move through placement states, carry care and medical histories, and stay linked to relevant conversations. I designed an interface around that model that feels natural to both coordinators and foster caregivers. I tested and refined the workflows through usability reviews with the rescue team, foster volunteer test users, designer colleagues, and UX advisors.

FOSTYENGINEERING

A platform animal rescues can trust

Building software for a cause I care about means its quality reflects my values, not just my skills. I architected the full-stack application around two roles: coordinators manage animal records and coordinate the entire placement pipeline, while volunteers see fostering opportunities and make requests.

Messaging is part of the coordination model rather than a separate inbox. Animal and group tags carry record context into each foster’s shared conversation with staff, keeping updates anchored to the care they support.

I designed reliability into the workflows rescue teams depend on. For request approvals, a transactional PostgreSQL function locks the request and updates assignment, status, and visibility records together. Row-level security and organization-scoped permissions isolate each rescue’s data. Fosty was engineered to be secure, reliable, and trusted.

TECHNICAL HIGHLIGHTS

Role-specific coordinator and foster workflows
Animal lifecycle, group and placement management
Transactional request-approval workflows with assignment conflict safeguards
Contextual, real-time messaging with animal and group tags
Organization-scoped permissions and row-level security

TECHNOLOGY

React
TypeScript
Vite
TanStack Query
React Router
Supabase
PostgreSQL

FOSTYOUTCOME

The future of Fosty

Providing relief to rescues everywhere.

I am actively rolling out Fosty with Colorado Kitty Coalition as my pilot partner. My primary objective is to ensure Fosty works reliably in their day-to-day environment. The next benchmark is full adoption by CKC, with the platform tailored to fulfill their foster-management needs.

The broken system extends beyond CKC. Once Fosty is proven in a rescue environment, I plan to find more partners and reach out to rescues and shelters interested in Fosty. I believe Fosty can fill a missing piece in animal rescue management and that shelters deserve a dedicated platform for foster operations.

Every design decision has a real user on the other end caring for a living animal. My personal connection to the problem, the organization, and the animals means I’m never satisfied. I am building an app that I would use myself. I am devoted to making product and technical decisions that improve the lives of the animals I love.

Fosty represents who I am as a designer and engineer. I like to build software close to users, grounded in real operations, and engineered for trust in high-stakes environments.

If you want to help with the Fosty mission or follow the journey, please reach out. There is always a demand for foster volunteers, so if you are interested in fostering, connect with your local animal shelter. I made Fosty to serve a community that I am passionate about, and I hope I inspire others to explore projects that make the world a better place.

CRUX VISIONORIGIN

PRODUCT DESIGNER · FULL-STACK ENGINEER

Crux Vision

2025 TO PRESENT

Crux Vision is a movement-review tool I created to turn climbing footage into a workspace for examining motion and technique. It uses pose data to create video overlays that reveal new layers of visual information, helping climbers examine movement holistically and magnify subtleties that ordinary playback can obscure.

Crux Vision introduction reading Find the move that matters and See your climbing in motion.

Climbers are always trying to improve, whether we are building strength, refining our technique, or working to complete a route at the edge of our ability. We are constantly looking to learn and grow, which makes analysis, technique, and movement comprehension central to the sport. It is also one reason climbing is so social: we learn from one another and work together to overcome challenges.

While climbing, you experience everything from a first-person perspective. You are limited to what you can see and feel, and you cannot observe your own body completely. Sometimes that means missing the bigger picture: how your body moves through the full sequence of a route. Watching another climber, or reviewing footage of yourself, reveals details and relationships that are difficult to recognize while you are on the wall.

Video extends that collaborative process by giving climbers an outside perspective they can revisit. Crux Vision complements video investigation by creating a container for deeper analysis. Movement trails can reveal the path of the hips, the timing of a foot swing, the arc of a reach, or the relationship between different parts of the body.

Using Crux Vision, I have experienced the satisfaction of seeing technical theory represented visually. It has confirmed ideas I had about why a move was or was not working, exposed relationships I had missed, and directed my attention toward things I would not have thought to examine.

Crux Vision cannot replace the intuition and experience of a climber. It complements the practice of analyzing movement and technique. It is a microscope for video analysis: a way to investigate movement, magnify subtleties, and build stronger connections between what we see and what we feel on the wall. The current public beta is only the beginning, with significant room to explore new visualizations, comparisons, and meaningful measurements.

CRUX VISIONMOVEMENT REVIEW

Review the crux

Crux Vision combines focused pose analysis with precision playback controls, making it easier to isolate a move and investigate a specific question. Select the short segment that contains the movement you want to understand, then move fluidly between ordinary video review and pose-based overlays.

ISOLATE THE CRUX

Select the section that contains the move you want to understand, often the hardest move, a recurring fall, or a place where two methods differ. A focused range keeps analysis centered on relevant climbing and lets the on-device pose model return results much faster.

REVIEW WITH PRECISION
Loop the selected range, slow down playback, step through analyzed frames, and create named checkpoints for important positions. Repeat a movement, inspect individual moments, and return quickly to the parts of an attempt that deserve closer attention.
FOCUS THE INVESTIGATION
Choose the body part that matches the question you are asking. Focus on an ankle generating momentum, or examine a hip, shoulder, wrist, knee, or elbow. Each selected trail turns a movement path obscured through time into something visible. This can help confirm an observation, reveal a new detail, or explain the movement to another climber.

CRUX VISIONVISUAL OVERLAY

Movement made visible

Movement trails trace parts of your body through space, preserving the complete shape of a movement as a persistent visual path. For this comparison, I selected my left ankle, hip midpoint, and shoulder midpoint to investigate two attempts at the same dynamic move.

THE MOVE

The move begins from a poor foothold: a flat, sideways-facing surface that is difficult to jump from without slipping. The destination holds are two opposing side-pulls. To stay on the wall, I need to catch them with enough height to keep my arms bent, create compression through my upper body, and press my foot into the flat wall.

It is a quick, coordinated movement, and the momentum, body position, and timing all have to come together within a fraction of a second.

READING THE TRAILS

In the unsuccessful attempt, I try to power directly through the move by pulling on the sloped handholds and jumping. My chest and shoulders lead while my hips follow behind. I arrive with my arms extended and my body stretched away from the wall, without enough leverage or any support from my feet to hold the position.

In the successful attempt, I generate momentum differently. I swing my left leg backward, then drive it forward and upward. The ankle trail makes that larger arc immediately visible. The swing carries my hips through the movement, allowing them to lead rather than trail behind my shoulders. The hip and shoulder trajectories rise higher, and I arrive more centered beneath the holds, with bent arms and a foot pressing into the wall to take weight off my arms.

My climbing experience gave me a theory about why one attempt worked and the other did not. Crux Vision gave that theory a visible form. Isolating the move and comparing trails pulled attention toward the timing and relationships that ordinary playback spreads across a fraction of a second.

The overlay does not replace the experience of a climber. It complements that knowledge, providing an intuitive surface for individuals, groups, and coaches to examine movement, make nuanced observations, confirm theories, and explain complex technique to one another.

CRUX VISIONENGINEERING

Building visuals from video

Drawing a skeleton is straightforward compared with deciding when its data is trustworthy enough to show. Climbing gives a pose model difficult input: crossed limbs, occlusion, motion blur, and unusual body positions the model was not trained for. Sometimes these conditions result in MediaPipe output with missing data or detection errors. The engineering challenge was to preserve the movement without disguising that uncertainty.

TECHNICAL HIGHLIGHTS

Progressive, on-device pose analysis in a module worker
Presentation-timestamp synchronization for live overlays
Immutable raw pose data with derived, inspectable views
Confidence-aware filtering and gap-bounded smoothing

TECHNOLOGY

React
TypeScript
Vite
MediaPipe Pose
MediaBunny
Canvas 2D
Web Workers

FROM VIDEO TO OVERLAY

A climber imports a local video and selects only the range they want to study, keeping the analysis focused and reducing processing time. MediaPipe analyzes that range progressively in a worker, drawing the overlay as results arrive so the climber can begin reviewing the movement while the source video remains playable. Each pose sample keeps its presentation timestamp, and live Canvas layers use the same display transform as the video, keeping overlays aligned across portrait and landscape footage without uploading or re-encoding the clip.

PRESERVING UNCERTAINTY

Raw pose results remain immutable. Confidence, body-scale plausibility, and motion over time determine which joints enter a derived view. Implausible jumps and unreliable positions are rejected; smoothing stays inside valid segments. When a joint disappears behind the climber or wall, Crux Vision shows an honest gap instead of inventing a continuous path.

CONTINUITY IS A TRADEOFF

The ordinary interface reduces a large calibration system to three understandable approaches. More continuous data can be useful, but it is not automatically more accurate.

BALANCEDUses moderate confidence and motion cutoffs to balance useful continuity against false positions.
STRICTUses higher confidence cutoffs and tighter motion limits, creating more gaps to avoid false limbs.
PERMISSIVEUses lower confidence cutoffs and looser motion limits to preserve more data, with a greater risk of questionable positions.

CALIBRATION BY ITERATION

Calibration is an iterative process rather than a search for one universally correct filter. I compare raw, accepted, rejected, One Euro, and centered views against the same cached MediaPipe results, isolating each policy or filter change from a new inference run. Coverage, rejection, gap, and smoothing metrics reveal how a setting changes the data, but visual review determines whether the resulting body positions still make sense.

The public beta exposes this deeper workspace for advanced manual calibration. Thresholds, continuity rules, smoothing behavior, and preview modes can be adjusted while the original pose samples remain immutable. Most climbers only need the Balanced, Strict, and Permissive presets; the advanced controls make the reasoning behind those presets inspectable and give me a controlled surface for developing better defaults.

Calibration remains ongoing. I am continuing to test across static positions, explosive dynamic moves, overhangs and occlusion, crossed limbs, varied camera angles, and different styles of climbing. A change should improve more than one kind of movement without creating a new failure somewhere else.

Crux Vision’s advanced pose-quality calibration workspace showing preview controls, quality metrics, and body-group coverage.

SMOOTHING RECORDED MOVEMENT

Even accepted pose landmarks can shift slightly from frame to frame. That noise makes a skeleton appear to jitter and turns a movement trail into a jagged path. Smoothing produces more visually stable poses and cleaner overlays and trails, but every smoother must trade some immediate responsiveness for visual continuity.

The One Euro filter is adaptive and causal: it uses the current and earlier accepted samples, applying more smoothing to slow or nearly still movement and responding more quickly as movement speed increases. Because it does not look ahead, it follows the chronology of the incoming pose track, but it can still visibly trail a fast, dynamic move.

The centered offline smoother uses accepted samples before and after each timestamp. Its symmetric window avoids the same systematic trailing and produced the most useful display for recorded dynamic movement while still removing substantial jitter, so it became the display default. Looking forward creates a different risk: movement can appear to begin slightly early. Both smoothers stay inside accepted segments and reset at missing or rejected data rather than blending across a gap. Accepted raw retains the original model output.

Crux Vision’s global confidence and body-group override controls for visibility and presence thresholds.

CONFIDENCE BY SCOPE

MediaPipe attaches two confidence signals to each detected joint. Visibilityestimates whether the joint is clearly visible rather than hidden by the body, wall, or another limb. Presence estimates whether the joint is actually within the captured frame. Crux Vision requires both signals to clear their configured thresholds before a joint enters an accepted view.

A global threshold establishes the baseline, while body-group overrides can target repeated problems in related joints without over-filtering the entire body. Joint-level overrides provide an even narrower adjustment when needed. Higher thresholds remove more questionable positions, reducing false limbs at the cost of more missing data and broken trails. Lower thresholds preserve more continuous movement, but increase the chance that an uncertain or incorrect position will remain visible.

Crux Vision’s advanced controls for confidence hysteresis, temporal plausibility checks, One Euro smoothing, and centered smoothing radius.

CONTINUITY AND PLAUSIBILITY

Confidence hysteresis uses different requirements for acquiring and retaining a joint, preventing borderline data from blinking on and off at a single cutoff. Timestamp-aware plausibility checks compare joint speed, acceleration, and changes in apparent limb length against body scale, rejecting positions that would require an implausible jump even when the model reports high confidence. The smoothing controls then tune filter responsiveness within the accepted values.

LIMITATIONS

Filtering and smoothing can reject noise or make accepted movement easier to read, but they cannot recover a joint the model never observed. One Euro can lag fast movement, while centered smoothing can anticipate movement onset. Both remain approximations that need continued comparison against accepted raw data.

Movement trails are most useful with a fixed or nearly fixed camera. Camera movement can distort trails because the overlay measures motion within the image, including movement introduced by panning, zooming, or camera shake.

Computer vision remains imperfect. Calibration can improve the usefulness of the visual result, but it cannot repair major detection errors or missing data. Crux Vision should not be treated as flawless motion capture or biomechanical truth.

CRUX VISIONOUTLOOK

An open investigation

Crux Vision began with climbing because it is the movement discipline I am familiar with and passionate about. Other sports and movement disciplines could benefit from the same concept: visuals can create a shared language between what someone feels, what a video shows, and what another person can observe.

The public beta is the first useful version of that idea. My immediate goal is to make the complete workflow comfortable during an ordinary gym session, then let real use guide what comes next. I want future features to grow from the questions climbers and coaches bring to the tool, not from adding visual complexity for its own sake.

NEW VISUAL LENSES

There is substantial room to deepen the investigation. New overlays could show the direction and speed of a movement, preserve earlier positions as ghost poses, or bring several settings together around a question such as hip drive, a leg swing, or the sequence of body positions through a move. Carefully bounded measurements could help examine timing, angles, stillness, and movement paths while remaining honest about missing or uncertain pose data.

COMPARE ATTEMPTS

Video comparison is especially valuable. Synchronizing two attempts could make differences in path, timing, body position, and method easier to see. Over time, saved review sessions, editable annotations, and shareable visuals could turn Crux Vision into a richer surface for collaboration between climbers and coaches. The same principles may complement expertise in other movement disciplines, but each brings its own questions. I would want those directions to be explored with the people who understand them.

Crux Vision will never pretend to know the correct way to move. Its purpose is to help athletes learn something about their technique, give coaches a clearer way to explain an observation, or add a new element to an open conversation between climbers.

Try Crux Vision with your own climbing video. If it helps you notice something—or if you have observations or improvements—I would love to hear about it. I welcome feedback, feature requests, and open conversations about where the project should go next.

INHERITANCEEXPERIENCE

MACHINE LEARNING & MOTION CAPTURE ENGINEERING INTERN

Inheritance

FALL 2025

Motion Data for Machine Learning

At Inheritance, I built a Python and Blender pipeline that converted the entire AMASS research dataset into standardized skeletal animation for machine learning training in Unreal Engine. The pipeline processed 11,265 motions from 344 subjects, representing more than 65 hours of motion capture. The team estimated that around 90% could support training, saving roughly three months of data generation and $70,000 in equivalent motion capture production.

11,265: NEW MOTIONS AVAILABLE
3 MONTHS: DATA GENERATION SAVED
$70K: MOTION CAPTURE VALUE

A sample of retargeted motion capture animations.

During my internship, Inheritance operated under the name KikiTora. The company turned ordinary video into detailed human motion, with a longer-term goal of making recorded behavior available as structured data for machine learning.

The training workflow ran in reverse: known motion drove standardized characters in Unreal, where controlled camera, lighting, and rendering variations created many video-to-motion pairs. Compatible motion capture was the limiting factor, which made this pipeline especially valuable.

INHERITANCECHALLENGE

The Motion Data Bottleneck

Training a model to generate motion capture from video requires matched examples of both: a video input and the exact 3D motion behind it. Producing enough of these pairs at the scale required for machine learning is difficult and expensive.

Collecting video and motion pairs directly was not practical at the scale the model required. Instead, we started with known motion capture and used it to generate the matching video. The animations drove standardized characters in Unreal Engine environments, where the same performance could be rendered from different camera angles with different lighting, character meshes, image quality, and levels of occlusion. Every render remained paired with the original animation, giving the model both its visual input and the correct 3D motion.

This may seem backwards when our goal is to turn real video into motion capture. However, starting from motion solved the problem of obtaining an exact label for every frame. It also made video variation effectively unlimited. One performance could produce countless training examples, but every example still depended on the same underlying movement. The number and diversity of available motion capture performances remained the main constraint.

At the time, Inheritance produced much of its own training motion capture. This was a costly operation requiring a capture stage, performers, and production staff. Existing animation data is difficult to license for machine learning, while new capture sessions are expensive and time-consuming. Expanding the motion library meant either producing more performances or finding another source the company could use.

AMASS, short for Archive of Motion Capture as Surface Shapes, offered another source. It is a large research archive that combines motion capture from many academic datasets into a shared human-body format. The archive contained more than 65 hours of motion across 11,265 animations and 344 subjects.

The as Surface Shapes part of the name captures the compatibility problem. AMASS represented motion through SMPL-H body-model parameters that produce a rigged human surface mesh. Inheritance needed conventional animation keyed onto its exact production skeleton, ready to move through Blender, GLB, and Unreal. In other words, the source was organized around a deformable body surface and its parameters, while the destination was organized around animation keys on a specific skeleton. The motion had to be reconstructed in 3D, aligned with the target hierarchy, rest pose, proportions, and coordinate spaces, and exported in the required format.

A wide collection of AMASS body models showing varied poses, movements, and body shapes. — THE MOTION AND BODY DIVERSITY REPRESENTED IN AMASS.

INHERITANCEENGINEERING

Rebuilding Motion in 3D

AMASS offers thousands of motions, but making them usable required a complex engineering solution. I built an automated Python and Blender pipeline that reconstructed each sequence from SMPL-H parameters, transferred it onto Inheritance's fixed production armature, and exported consistent GLB animations for Unreal Engine.

Ordinary retargeting starts with two complete skeletons. Here, the source files did not contain a conventional keyed armature. They stored joint rotations and root movement inside a parametric human-body mesh model. The pipeline first had to recover where every joint moved in three dimensions, then turn those trajectories into animation on a skeleton with a different rest pose, proportions, hierarchy, and coordinate spaces.

AMASS MOTION, REBUILT ON A PRODUCTION ARMATURE.

TECHNOLOGY

Python
Blender
NumPy
SMPL-H
glTF / GLB
Unreal Engine

CORE PIPELINE

Each frame contained rotations for 52 SMPL-H joints plus the movement of the root joint. I used forward kinematics to reconstruct the body's world-space joint positions through the parent-child hierarchy. Those positions then drove Inheritance's standardized armature in Blender, where the evaluated motion was baked into keyframes and exported as an animated GLB.

The pipeline for each animation:

Read the joint rotations, root movement, and source frame rate.
Reconstructed every joint's position through the SMPL-H hierarchy.
Aligned the motion with the target armature and coordinate system.
Drove and baked the standardized skeleton in Blender.
Exported consistent animated GLBs for the Unreal ML workflow.

ENGINEERING HIGHLIGHTS

RECONCILING REST POSES

AMASS baked in the SMPL-H template T-pose, while the production pipeline depended on a standardized base A-pose for calibration, mesh weight painting, and other downstream character setup. I calculated the transforms required to convert the motion from the source pose into the production pose so every animation used the same armature without changing the intended motion.

BALANCING PROPORTIONS AND MOTION

The reconstructed joint paths came from parametric bodies whose proportions did not match the production rig. Forcing every target bone to reach those positions exactly caused the skeleton to stretch, preserving the coordinates but distorting the body and the character of the movement. The solve was to treat the joint paths as motion guides rather than literal bone dimensions, then use Blender constraints and baking to transfer their timing and trajectories onto the standardized armature. This preserved the production skeleton's proportions while retaining the recognizable intent of each performance.

TRANSLATING THREE ROTATION SYSTEMS

The motion had to pass through three mathematical representations because each one served a different part of the pipeline. AMASS stored each joint rotation in axis-angle form, a compact three-value encoding suited to a large dataset but not to composing an entire skeletal hierarchy. I converted those values into rotation matrices with Rodrigues' formula so parent and child transforms could be combined through forward kinematics and evaluated across local and world coordinate spaces. The baked animation ultimately entered Blender and GLB's quaternion-based representation, which supports stable rotation and interpolation in animation tools.

Although all three representations encode orientation, they are not interchangeable without conversion. Every rotation had to remain in the correct coordinate frame as it moved from compact source data, through hierarchical calculations, and into the exported animation. Errors could turn joints around the wrong axis, reverse a limb, or allow small errors to accumulate through the hierarchy.

INHERITANCEIMPACT

Accelerating the Data Roadmap

The AMASS retargeting pipeline gave Inheritance a large, usable motion library without requiring equivalent new motion capture sessions. It processed all 11,265 animations available to the company, with approximately 90% qualifying as useful for training. The pipeline provided thousands of performances across 344 subjects and greatly expanded the team's capacity to generate paired video and motion for training.

Astrid Wilde, CEO of Inheritance, estimated that the pipeline accelerated the company's data-generation roadmap by approximately three months and avoided roughly $70,000 in equivalent stage, performer, and production costs. It also broadened the foundation of human motion available for controlled training renders.

FROM MOTION CAPTURE TO PHYSICAL AI

At KikiTora, the immediate goal was to recover detailed human motion from ordinary video. Inheritance now frames that idea more broadly, converting video into structured representations of behavior, contact, and action that robots and world models can use for training. The opportunity extends beyond animation. For computer vision, it means moving from recognizing what appears in a frame toward recovering how people move and interact with the physical world over time.

At scale, that approach could allow robotics teams to draw from recorded human behavior instead of recreating every demonstration for each task, environment, or machine. A growing body of structured examples could help models generalize across a wider range of conditions.

TURNING RESEARCH INTO IMPACT

Developing the AMASS retargeting pipeline was especially rewarding because it tangibly accelerated the mission by introducing thousands of motions into the ML training workflow. My work addressed a serious production constraint and gave new life to data beyond its original research setting.

I was able to draw from several areas of my experience, from data infrastructure and 3D math to computer vision and Blender automation. The project brought together model training, computer graphics, and animation production, and I enjoyed figuring out how to make those systems work together. The result was a reliable pipeline that delivered measurable value to Inheritance while pointing to a much larger opportunity for structured motion data across computer vision, animation, robotics, and physical AI.

VALEXPERIENCE

FULL-STACK ENGINEER (CONTRACT)

Val

MARCH 2026 TO JUNE 2026

Owning a recovery platform from product decisions to production releases

At Val, I shaped and shipped a live recovery and post-discharge platform used by behavioral-health providers and people continuing their care. My work spanned product decisions, full-stack development, quality assurance, and production releases. I joined while the product was still early but already in production with a treatment provider and real users and that combination required that I deliver immediate impact. The platform had an important mission, active use, and the usability and reliability problems that come with software developing quickly.

I worked closely with the founders and technical leadership in a small early-stage team. I was given broad ownership over what to investigate and freedom to deliver solutions with the highest impact. I evaluated which problems were most important, scoped solutions, worked across the web application, backend, and database, and carried changes through release. Priorities came from user feedback, the workflows people depended on most, requests from the team, and opportunities I identified through my own investigation.

I was drawn to Val because it offered the responsibility and freedom of early-stage product development and a mission that inspired me. Technology plays an important role in healthcare, but it has not always been designed around the people who rely on it. Continuing care after treatment is a difficult transition, and software should make support easier to reach rather than introduce another barrier. At Val, improving usability and reliability was connected to a larger purpose: helping people remain engaged with their care and recovery communities.

This experience is a strong demonstration of how I work as a software engineer. I collaborated closely with a team while exercising independent judgment, owned features from idea through release, moved across the full stack, and took responsibility for delivering reliable software to production.

VALCONTRIBUTIONS

Product decisions and full-stack delivery

Working on an early product requires moving between new features, user experience, production issues, and technical foundations. User feedback showed us where people were struggling, and as I worked across the application, I uncovered additional problems and opportunities. I had the freedom to investigate those needs, propose solutions, and focus on the work that would make Val more useful and dependable.

PRODUCT JUDGMENT

I prioritized work through user feedback, product usage, founder input, and direct investigation of the application. Some needs were urgent production problems while others were workflows that had outgrown their first implementation. I learned to balance immediate fixes with improvements that would remain useful as the product developed.

USER EXPERIENCE

I improved experiences across both sides of the platform: the application used by people continuing their recovery and the tools used by treatment providers and care partners. My work touched dashboards, journaling, daily reflections, announcements, client management, and alumni-to-coach calling workflows. I removed friction and made the product more usable across user roles, devices and configurations.

FULL-STACK DELIVERY

Owning a feature meant owning the system behind it. A change that began in the interface could require new data models, backend logic, permissions, migrations, or release changes before it worked reliably for users. I followed those dependencies across the stack and carried the work through production. That continuity allowed me to deliver complete solutions and fix the cause of a problem, not just its most visible symptom.

FEATURE SPOTLIGHT

Bringing meeting discovery into Val

One of the clearest examples of my product ownership was Val's AA and NA meeting finder. The existing experience sent users to a third-party website through a built-in browser. It technically offered access to meeting information, but it was fragmented, inconsistent, and not a strong long-term experience for a product intended to support continuing care.

I researched the public meeting-data ecosystem and proposed bringing discovery directly into Val. The challenge was that meeting information was spread across regional organizations using different feeds, fields, formats, and levels of completeness. I designed and built a full-stack system that aggregated more than 5,500 meetings from 19 regional AA and NA sources and normalized them into one consistent experience. The system was built to grow with Val, allowing new AA intergroups and NA service bodies to be added as the platform expanded to more locations and treatment partners.

Users could search meetings and filter by day, time, attendance format, and distance. Location-aware sorting helped them find relevant options nearby, while online and hybrid meetings remained available when geography was not the deciding factor. Supabase Edge Functions fetched the regional data, a shared model reconciled the source formats, and a layered caching and scheduled warming process kept the experience responsive without placing unnecessary load on upstream services.

The meeting finder began as an opportunity I identified independently and became one of the product's most valued features. Val's founders informed me it was a meaningful draw for the platform, and users reported relying on it daily to find meetings in their area and support their continuing aftercare. It is an explicit testiment of the broader contribution I made at Val. I am a high agency, high impact engineer who can carry a better experience across research, product decisions, engineering, and production.

VALPRODUCTION

Carrying changes safely into production

I regularly carried changes from development through production across Val's web application, backend services, and database. The work required understanding the dependencies between application code, customer configuration, data, and infrastructure. It also required evaluating release risk and responding when one part of the process failed.

QUALITY ASSURANCE

Building a regression-testing foundation

Val was moving quickly without enough automated protection around its core patient and provider workflows. I established an end-to-end regression suite in Playwright and integrated it with GitLab CI against a containerized Supabase environment.

I made the tests repeatable by creating seeded application states and consistent local and CI environments. The automated environment also exposed authentication timing and loading-state problems that I traced and resolved. The suite gave us dependable checks for important workflows before production and clearer diagnostics when a test failed.

RELIABILITY AND DATA ACCESS

Production reliability required work beyond the visible interface. I corrected asset caching that could leave users on stale application versions after a release. In PostgreSQL, I fixed Row-Level Security policies, grants, and ownership issues that blocked legitimate workflows or caused recursive queries.

PRODUCTION RELEASES

Val's releases crossed customer-specific web builds, database changes, and cloud services. I regularly coordinated builds, pending Supabase and PostgreSQL migrations, Git-SHA-versioned backend containers, front-end publishing through AWS, and final verification.

Alongside this work, I consolidated staging and release procedures, strengthened migration handling, and turned recurring operational steps into reusable tooling. This created a clearer release path with explicit checkpoints for halting, diagnosing, and recovering when something failed.

TECHNOLOGY

React
TypeScript
Next.js
Capacitor
TanStack Query
Supabase
PostgreSQL
Playwright
GitLab CI
Docker
Nx
Terraform
AWS