Research Log
Notes, demos, and deep dives
Follow along as we build. Lab experiments, engineering breakdowns, and lessons from the frontier.
We Built AI Employees — Not Assistants, Actual Employees
We built AI employees with job titles, responsibilities, and performance reviews. Then we built mission control to manage them.
40 Seconds Was Too Slow — Making Yield Faster and More Reliable
40 seconds per turn was too slow. Eman parallelized agents, fixed caching, and cut latency by 12%. Here's how.
75% Pain Points, 10% Everything Else — Fixing Coverage Imbalance
Pain points had 75% coverage. Everything else? Under 10%. Mugi shows how we taught Yield to go wider, not just deeper.
Our AI Stopped Responding — Fixing Silent Failures in Yield
Sometimes our AI just stopped responding. Mugi shows how we fixed silent failures and made Yield reliable.
Power Follows Visibility — Introducing Yield
Yield interviews every employee, connects every system, and builds a live map of how your company actually works.
"I Like Jira" Became "Has Issues With Jira" — Fixing AI Data Accuracy
"I like Jira" became "has issues with Jira." Zeineb shows how we're fixing data accuracy in Yield's AI interviews.
Why Trust Is the Biggest Challenge in AI Interviews (And How to Fix It)
Research-backed principles that make AI interviews feel more natural and trustworthy.
What Makes An AI Interviewer Trustworthy? (It's Not What You Think)
What makes an AI interviewer trustworthy? Not intelligence — behavior. Four research-backed principles for building Yield.
Making Controlled AI Videos - What Actually Works
Eman breaks down how to make controlled AI-generated videos — why consistency is hard and the strategy that actually works
Building a Sign Language Avatar That Actually Works
This avatar looked possessed. Then I stopped fighting the skeleton mismatch — and it worked immediately.
The Hardest Part of Medical AI — 3D Reconstruction
This 3D tumor visualization looks impressive, this system works. But more importantly, it fails loudly when it should.
Turn Any Video Into a 3D Space You Can Explore
Upload a video, get a 3D space you can explore, annotate, and measure. No scanners or technical skills required.
Building a Medical AI That Doesn't Lie
A brain tumor detection AI that shows its work, expresses uncertainty, and fails loudly when it's not confident.
Draw a Board Game on Paper, AI Makes It Playable
Draw a board game on paper, scan it, and watch AI turn it into a playable digital game with auto-generated rules.
Fine-Tuning, Medical LLMs, and Clinical Alignment
Zeineb improves AutoScanAI through fine-tuning and adds a Medical LLM Explainer to deliver robust, clinically interpretable, WHO-aware AI outputs.
Designing Medical AI Without Oversimplifying Reality
Zeineb explains the motivation and design behind AutoScanAI, a clinically grounded medical AI system built for early brain tumor detection using MRI data.
A Deeper Look at Gaussian Splatting and Its Pipeline
Kevin breaks down how Gaussian Splatting works for 3D reconstruction and explains the design choices behind a web platform built to make it easier to use.
How the Grid-Reading Pipeline Works for AI Board Games
Eman explains how his grid-reading pipeline works, showing how hand-drawn grids are processed into structured inputs for AI-generated board games.
Face Tracking and Fingerspelling for ASL Motion Generation
Mugi updates SignMate's motion generation system by adding face tracking and fingerspelling to improve ASL accuracy, expressiveness, and realism.
Building a Grid-Based Interface for AI-Generated Games
Eman demos a website that reads grid-based inputs and game elements, helping AI reliably understand board game layouts, mechanics, and rules.
Building an AI That Finds Brain Tumors in MRIs
Zeineb introduces her medical AI project focused on early brain tumor detection, explaining how AI, precision, and analysis can improve patient outcomes.
Exploring Apple's 3D Reconstruction Pipeline
This session breaks down Apple's 3D reconstruction approach, exploring its design choices and how the tools can be applied or extended in future projects.
Developing 3D Brain Models to Detect Tumors
Zeineb shares progress on her model, discussing recent experiments and how she's refining the system while evaluating results as the project moves forward.
Why a Dual Camera Setup Matters for Motion Generation
Mugi updates her project by moving from a single-camera setup to a dual-camera solution, improving motion capture accuracy and overall system reliability.
Updating the Camera and Robot Setup in Isaac Lab
Eman improves the camera and robot setup in Isaac Lab, building a more stable and accurate system for future autonomous movement.
Teaching a Robot to Seek Rewards Autonomously
Eman showcases a major robotics update, where the robot now navigates autonomously toward rewards with high accuracy based on its environment.
Robots, AI robots everywhere!
Check out Muhammad Eman Aftab's work from our Budapest research lab, where he builds AI projects in public, sharing progress, challenges, and learnings.
Comparing AI Agents for Image Analysis
Kevin compares different AI agents for image analysis, showing how each approaches visual tasks and where certain tools perform better by use case.
From Land to Digital World: Introducing Cenarius
Kevin introduces Cenarius, a project that scans real-world environments like farmlands and construction sites into interactive 3D spaces.
Camera Calibration and Isaac Setup Explained
Eman advances his vision-based system by upgrading camera calibration from 2D to 3D and setting up the Isaac environment for testing and future autonomy.
Venato, Simplifying Regulation Monitoring with AI
David showcases Venato, an AI-powered regulation monitoring platform, explaining how it uses AI to collect, structure, and surface regulatory information.
Exploring ASL Motion Capture with MediaPipe and RPM
Mugi shares early SignMate experiments capturing ASL motion, testing MediaPipe, RPM, and ASL Mocap datasets to evaluate gesture accuracy and complexity.
Exploring ASL Motion Capture with Plask & FreeMoCap
Mugi tests Plask and FreeMoCap for ASL motion capture, comparing how well they track gestures and outlining strengths, limits, and animation challenges.
Why Noise-Free Data Matters for ASL Avatars
Munkhchimeg Sergelen explains how SignMate avatars are created and why clean, precise motion data is essential for accurate ASL handshapes and movements.
Building Motion Generation for Sign Language AI
Mugi explains how SignMate converts ASL gloss into 3D avatar motion, walking through the system that maps language inputs to realistic signing gestures.
How ASL Structure Shapes Our SignMate Model
Munkhchimeg Sergelen explains how ASL differs from spoken English and why these differences are essential when designing SignMate.
Boosting Tumor Detection Accuracy
Zeineb explains why she upgraded her MRI tumor-detection model from ResNet18 to ResNet50 to build a stronger system for more accurate detection.
Eman's Big Idea at the Lightbloom Lab
Eman introduces his Lightbloom AI Lab project: a system designed to turn simple hand-drawn sketches into fully playable board games.
Real-Time 3D Reconstruction with Gaussian Splatting
Kevin demonstrates real-time 3D reconstruction with Gaussian Splatting — high-quality scans without expensive hardware or compute.
Vision-Driven Robot Movement
Eman demos a vision-driven robot control system that reads printed signs to interpret commands and move autonomously in real time.
Experimenting With a Real-Time Sign Language Avatar
Inside the Lightbloom Lab in Budapest, Mugi walks through her latest experiment turning teammate Kevin into a real-time signing avatar.
How We Run Workloads on Our New Computer Cluster
Kevin demonstrates how to use the Lightbloom Lab's newly assembled computer cluster, the shared infrastructure powering all ongoing projects.
AI That Turns Your Sketches Into Playable Board Games
Join us inside the Lightbloom AI Lab in Budapest as our researcher, Eman, demonstrates the latest updates to our Sketch-to-Game Generator.
AI for Reliable Brain Tumor Detection
Zeineb demonstrates the latest improvements to the Autoscan AI model, advancing early brain tumor prediction at the Lightbloom AI Lab in Budapest.
How Mugi Is Building SignMate
Meet Mugi, an AI researcher in our Budapest Lab, building SignMate to make communication more accessible through speech-to-sign avatars.