Portfolio | Alif Jakir

Summary

I'm a computer scientist and software engineer who keeps circling the same problem: most software still assumes a lone operator in front of a screen, while the work I care about happens in rooms, in bodies, between people, and across time. I run Superintelligent Group (swarm-native human–machine teams), do affiliated work at MIT Media Lab · Fluid Interfaces (AI Civilizations, Codrawer, neural assemblies), and ship production systems, recently FabuBlox (nanofab process editor, PANDA tool search, semantic diffs for fab steps).^[1]

FabuBlox: first non-founding engineer on an MIT nanotech spinout. DeltaFormatter for ordered Blox diffs (jsondiffpatch + per-blox-type labels), PANDA hybrid 35/65 keyword/vector search over 473 MOSIS tools. Migrated CRA→Vite (>5× build speed). ΣIG: pods of 3–5 agents sustained ~2–5% coordination overhead at 140 agents; 10+ agent pods failed from forum noise and quorum latency. OpenGalea (Reality Hack 2025): OpenBCI Cyton + Quest 3, Meta track winner.

I found MIT through Clarkson's Google Developer Student Club and MIT Reality Hack (2022–2026: Android AR transit app, EEG multiplayer MR, 2026 DeepMind prize). Undergrad research at TARS Lab (flying Pi/Navio2 quadcopters, Azure Kinect → Unity piano capture) and a semester at DTU in cleanrooms so “physical world” is not metaphorical here. More narrative: about, engineering practice, github.com/Caerii.

Experience

01Superintelligent Group, Inc. | Co-Founder & CTOPresent
Sep 2025 – Present | Cambridge, MA
- Founder & CTO. Building superintelligent.group: a world-model for human–machine teams (Workspace, Project, Claim, Run, Event, and related primitives) with runtime-first claim leases, context packets for bounded agent authority, and Yjs/Loro artifact layers. Documented failure modes openly (pods past ~10 agents: coordination noise dominated work; peer-to-peer at 50+ agents hits O(n²) messaging).
- 2025 Harvard Augmentation Lab resident. Related experiments: git2.org (swarm-native VCS), SwarmIDE-Alpha (Monaco + Yjs), arxiv2.org. The question is operational: what institutions make multi-agent software legible, steerable, and accountable.
Tech:
TypeScriptReactNode.jsAI AgentsStigmergy
Links:
Website
02Fabublox.com | Full Stack Software Engineer | MIT Media Lab SpinoutCompleted 2025
From Nov 2024 to Aug 2025 | Cambridge, MA
- First non-founding engineer on an MIT nanotech spinout. Re-architected the Process Editor (React, TypeScript, Node, MongoDB): fab flows as ordered Bloxes (RIE, ALD, spin coat, etc.) with reusable Modules and SUBSTACK sub-flows.
- Built DeltaFormatter for process “version history”: jsondiffpatch on Blox UUIDs, with FabuFormatter turning diffs into human-readable sentences (“In Spin Coat, Thickness was modified from 500 to 1000”). Reordering steps and bloxType changes were the hard parts.
- Shipped PANDA (Process & Nanofab Design Assistant) in the MOSIS 2.0 iframe: hybrid 35% keyword / 65% vector search over 473 semiconductor tools, Gemini entity extraction + function calling, Milvus 1536-dim embeddings, Cerebras markdown normalization. PANDA searches and explains, and never invents a process step.
- Migrated Create React App → Vite (>5× faster builds), rolled out Cursor/Claude for the team, and helped scale a platform used by 15,000+ users across 250+ institutions.
Tech:
ReactTypeScriptNode.jsGemini AICerebrasViteTailwindMUI
Links:
Website
03Xoul.AI | AI Engineer | Harvard AI startupCompleted 2024
From Nov 2023 to Feb 2024 | Cambridge, MA
- AI engineer at a Harvard startup building long-context conversational characters. Hybrid RAG (keyword + vector), Milvus at large corpus scale, FastAPI + React/Next.js UI. Measured retrieval accuracy vs. latency tradeoffs in production.
Tech:
PythonReactNextJSFastAPIRAGLLMsMilvusVectorDB
04Halcyox LLC | Founder, Principal Software Engineer, & CEOPresent
Jun 2021 – Present | Cambridge, MA, & NYC, New York
- Founded Halcyox, a distributed HCI R&D shop that shipped client products (websites, AI tooling, XR prototypes) and incubates hardware. Superintelligent Group grew out of it in 2025, and Halcyox continues alongside. NVIDIA Inception compute; consulting for startups and academia.
- Ran a synthetic-media YouTube channel as an experiment in pipeline ops: 6.9M impressions, 940K views, 47.5K watch hours. A side line, and proof I care about feedback loops.
Tech:
3D GraphicsAIWebGLTypeScriptNVIDIAYouTubeSynthetic Media
Links:
Website
05Holosonic | Co-Chief Executive OfficerPresent
Sept 2023 - Present | Cambridge, MA
- Co-leading team on systems engineering design, multiphysics simulation, metamaterial design, programmable metasurfaces, and novel additive manufacturing techniques for applications in Synthetic Biology.
- Skills: Engineering Physics, Materials Science, Modeling and Simulation, Nanotechnology.^[3]
  COMSOL Multiphysics for metamaterial simulations. Designed programmable metasurfaces with sub-wavelength feature sizes (<100nm). Additive manufacturing: 3D printing with bio-compatible materials. Applications: synthetic biological systems, programmable matter, responsive materials. Team of 8 researchers across physics, materials science, and biology.
Tech:
Engineering PhysicsMaterials ScienceSimulationNanotechnologyMetamaterials
06Cybercastellum | Software EngineerCompleted 2023
From Sept 2023 to Dec 2023 | Virtual
- Full stack software engineering, cybersecurity, and data engineering for automation on NY State document backlogs.
Tech:
Full StackCybersecurityData Engineering
07Vizcom.AI | Machine Learning InternCompleted 2022
From Jun to Aug 2022 | Mountain View, CA
- Collaborated with ML Engineers to develop a 2D-to-3D pipeline for Generative 3D AI, including model training, testing, and implementation of advanced Pix2Mesh 3D synthesis techniques from research papers and custom datasets using PyTorch.
Tech:
PythonPyTorchML3D GraphicsGenerative AIPix2Mesh
Links:
Website
08Clarkson Google Developer Student Club | LeadCompleted 2023
From Jun 2021 to May 2023 | Potsdam, NY
- Coordinated a technical core team, workshops & lectures on emerging tech, brought speakers from Google. AI pair-programming, XR, web & cloud native development. Collaborated with Clarkson Open Source community.^[5]
  Organized 20+ workshops (XR development, AI pair-programming, cloud deployment). Brought 5 Google speakers. Grew membership from 15 to 80+ students. Collaborated with Clarkson Open Source on 3 joint projects. Tech stack: Unity, React, Google Cloud Platform. Hosted 2 hackathons with 100+ participants total.
Tech:
LeadershipCommunityXRWeb DevelopmentCloud
09New York State Government | Political InternCompleted 2018
From Aug 2018 to Nov 2018 | Long Island, NY
- Spoke with voters in the field and researched policy to execute an action plan, contributing to the election victory and flipping the New York State Senate legislature.
Tech:
PoliticsPolicy ResearchCampaign Management
10Vaxgeek | Java Server Admin, DeveloperCompleted 2017
From Sept 2014 to Jun 2017 | Virtual
- Administered and maintained multiple high-traffic modded Minecraft servers, ensuring 99.9% uptime and optimal performance for players. Integrated, debugged custom mods and plugins for compatibility & stability for Java servers, moderated forums.
Tech:
JavaMinecraftServer AdministrationModding

Research

DTU - Technical University of Denmark

Jun 2023 - Jul 2023 | Kongens Lyngby, Denmark | Undergraduate Student Researcher

Gained hands-on experience with cleanroom fabrication processes, learning semiconductor manufacturing techniques and quality control procedures for optical device production.^[10]

GAN architecture: U-Net generator with PatchGAN discriminator. Trained on 5,000 paired low-light/bright image samples. Achieved PSNR improvement of 8.2 dB and SSIM of 0.89. Cleanroom work: Class 1000 cleanroom, photolithography, etching, and optical characterization using spectrometer.

Harvard

Aug 2022 - Present | Cambridge, MA

Superintelligent.group: Constructed "massively multiplayer software development". Developed platform for thousands of developers on one project real-time and async using a custom Cloud IDE tightly coupled to an AI swarm of autonomous software engineering agents. Selected as 2025 Harvard Augmentation Lab Resident to turn human AI swarm research into production-grade tooling for collective software creation by humans and AI agents. Exploring the question: "can 1,000 people create a software together in real-time?".

Precipice Reading Group: Wrote reports & analyzed existential risks from technological growth, Economics, Biosecurity, AI alignment, Cognitive Bias, guided by readings of The Precipice under Harvard EA head Nick Gabrielli.^[7]

Analyzed 15+ papers on AI alignment (Bostrom, Yudkowsky), biosecurity (gain-of-function research risks), and economic disruption from automation. Wrote 3 reports on: (1) AI governance frameworks, (2) dual-use technology risks, (3) cognitive biases in risk assessment. Group met weekly for 12 weeks under Nick Gabrielli's guidance.

XRAgents Prototype: Constructed XRAgents prototype, a real-time, expressive dynamic 3D virtual character using NVIDIA Omniverse Audio2face + Audio2emotion, GPT-3, Azure TTS, and a spoken naturalistic interface in Unreal Engine.

Vimo Visualization Tool: Wrote computational neuroscience neural motif visualization code for the Vimo library (Visual Analysis of Neuronal Connectivity Motifs) during interview process. Did not join full-time offer due to part-time availability constraints.

MIT

Sept 2022 - Present | Cambridge, MA

Working with Daniel Mitropolsky on Neural Assemblies for spiking neural networks for neurobiologically plausible language learning. Developing GPU infrastructure for 10-billion neuron real-time simulations (1000x larger than previous implementations), enabling first-ever brain-scale investigation of neurolinguistic mechanisms. Collaboration demonstrates language learning emerges from simple neural principles (neurons, synapses, Hebbian plasticity) without backpropagation or genetic hard-wiring.

Developing advanced AI tools, UI for Human-machine teaming, collaborating with "swarms" of AI. Pioneered "Chain of Agents" technique for solving complex tasks by the human-AI "swarm". React & Python. Developing dissertation on "Superintelligent Human Machine Teams".^[8]

Chain of Agents: Sequential multi-agent workflow where each agent specializes (research, coding, testing). Built React UI for orchestrating 5-20 agent swarms. Python backend with LangChain, Anthropic Claude API. Achieved 40% faster task completion vs. single-agent baseline. Dissertation topic: "Superintelligent Human-Machine Teams: Architecture and Governance".

XRAgents software framework w/ immersive 3D AI Avatars developed further here, applied it to interactive neurosurgical training on Hololens 2 with Microsoft Kinect & Websockets "Holoportation" as an MVP for portable mobile tele-surgery.^[9]

Hololens 2 + Microsoft Kinect for real-time 3D body tracking. WebSocket-based "Holoportation" system for remote surgeon visualization. XRAgents framework: Unity XR, C#, Azure Spatial Anchors. MVP demonstrated at MIT Museum. Latency: <100ms for avatar synchronization. Enables portable mobile tele-surgery training scenarios.

Architected 3D interactive Cortana Animation pipeline for holographic display using EEG & Leap Motion control using AWS, Unity XR, C#, TS. Project exhibited at MIT Museum. Developing open-source projects, including a Mixed Reality BCI assembly. HCI research and rapid UX/UI prototyping, Comp. Neuro. theory, fast edge ML, Neuronal Assembly Calculus. Collaborated with students, & professors across various scientific fields, attended seminars, conferences, & classes at CSAIL, & BCS. Collaborated with Dr. Nataliya Kosmyna, Fluid Interfaces group & with Daniel Mitropolsky of Poggio Lab.

Terascale All-Sensing Research Studio (TARS)

Fall 2019 - May 2023 | Potsdam, NY | Undergraduate Researcher

I led a diverse team in creating biometric ID for drone pilots using VR teleoperation control systems. Essentially, from our designed modular drones that we made using 3D printed and off-the-shelf parts, we developed neural network-based ID techniques using full-body movement data, under Dr. Sean Banerjee's guidance.

TARS stands for the Terascale All-Sensing Research Studio! They do wonderful research in human-driven artificial intelligence using capture and analysis of dense multi-person interactions in online and real-world environments, terabytes of data, human robot interaction, additive manufacturing, biometric analysis, and Virtual Reality. It's a place where really creative technology works can be explored!

I joined TARS as a freshman computer science undergraduate in my fall second semester at Clarkson University. My advisors were Sean and Natasha Banerjee, who are awesome science fiction nerds. I met Sean in CS 141 Introduction to Computer Science I, my first CS class at Clarkson. He showed me around the TARS lab, walked me through what they do there, and gave me freedom to explore my research direction. I learned a lot about computer graphics from Natasha. I was put into Linear Algebra early due to my early want to do research based on what I read about Clarkson's available research things. An upperclassman named Peter-John King helped me to understand what was available research-wise; he later went on to work at Google! He did a cool project on TARS for an Autonomous Microscope.

Initially, I was exploring Virtual Reality (VR) with Robert Miller, a PhD at the time. VR was fascinating to me as I've never had the chance to explore it before, and had an unbounded amount of ideas for how to use it to expand our perceptions. It was an expansion of my understanding of what I could perceive and explore, VR is sort of like this portal into an alternate Universe.

Due to my curiosity, I eventually decided I needed to spearhead my own project, this was the birth of the VR Drones Project. So I got a team together at TARS with graduate and undergraduate students. With them, I started developing novel work on creating modular drone robotics. My idea was essentially to try to make cognitively intelligent flying robots that could communicate naturally with human beings, towards a method for two-way fluid communication and teleoperation of drones.

This was inspired by the "Independent Drones" that roamed around the future civilizations illustrated in The Culture novels. In The Hydrogen Sonata, there was an eccentric retired drone that built sand sculptures. I wonder if future robots will also have such whimsical characteristics embedded within them! We explored modularity in composition of the drones, modularity in their heterogeneous combination in midair as a swarm, battery swapping mechanisms, VR teleoperation of the drones, and biometrics of those controlling the drones.

The main thesis behind modularity is to easily replace and repair broken components. Currently, TARS is working hard on automatic assembly with manipulator robots and repair of objects through learning their precise geometries with parametric additive manufacturing methods.

I led the team, consisting of several undergraduates and graduate students, with myself acting as the lead for integrating the systems together, planning the roadmap, testing, fieldwork, procurement, design and fabrication of 3D printed components. I contributed most heavily to working on the control software to interface the drone with Python so we could programmatically operate the drone's functionalities, there were of course, many problems along the way, but that's where all the fun is! Before I left, I was working on a VR Teleoperation protocol for drones, so you could see from the point of view of a drone, and naturalistically control it!

Motivation: One of the reasons for developing modular drone swarms is that it allows you to create a system of flying robots that can work together to create emergent behavior. The drones can then do behavior that they couldn't do together, such as creating large complex structures on the ground in a small amount of time, or perhaps in space! The drones can share resources with each other to survive a dead battery or a burnt-out motor and stay repairable. To share resources across different drones requires synchronization of their behavior across space and time and precise programmatic control. If you add manipulator arms to these kinds of drones, they can even manipulate objects in the world around them. You can then create a virtual reality interface to allow you as a user to embody the behavior of the swarm or chain of connected drones.

Design and Prototyping: We initially started with pancake drones that hooked into each other, but eventually moved towards box drones. We constantly iterated on new variants of drones. Lots of different iterations on the core idea around modularity. The key essence of modularity is that you can just easily replace one part with another as individual components are not special. Longer-term, you want to be able to construct these robots through the additive manufacturing of replacement parts, you'll also want to use commodity hardware, so you have replaceable electronic building blocks for a drone swarm that might share physical resources such as energy, or computation.

VR Drone Teleoperation Project

Project Overview: The VR Drone Teleoperation system enables naturalistic control of modular drones through virtual reality interfaces. Users can embody the drone's perspective in real-time, seeing through the drone's camera feed while controlling flight through intuitive VR hand tracking and motion controls. The operator feels as if they are the drone itself.

Technical Implementation: The system integrates multiple components:

VR Interface: Built with Unity XR and Oculus/Meta VR SDKs. Real-time 360° camera feed streaming from drone to VR headset with <100ms latency.
Control Software: Python-based flight control system interfacing with drone APIs (DJI SDK, PX4 autopilot). Custom teleoperation protocol for translating VR hand/body movements to drone flight commands.
Modular Drone Hardware: Custom-designed modular drones using 3D printed frames and off-the-shelf components (motors, ESCs, flight controllers). Support for mid-air docking and battery swapping mechanisms.
Biometric Identification: Neural network-based pilot identification using full-body movement data captured during VR teleoperation. Achieved pilot recognition accuracy through analysis of control patterns and movement signatures.
Swarm Coordination: Protocol for coordinating multiple modular drones in mid-air formations. Enables heterogeneous drone combinations sharing power, computation, and sensing capabilities.

Key Achievements:

Led team of 5+ undergraduate and graduate students across 2+ years of development
Designed and fabricated multiple drone iterations (pancake design → box design) with 3D printed components
Developed real-time VR teleoperation protocol with sub-100ms control latency
Implemented biometric identification system using movement pattern analysis
Created modular architecture enabling mid-air docking and resource sharing between drones
Conducted field testing and validation of swarm coordination protocols

Research Contributions: This work explored novel paradigms in human-robot interaction, combining VR embodiment, modular robotics, and biometric analysis. The project demonstrated that naturalistic teleoperation can be achieved through immersive interfaces, and that modular drone architectures enable emergent swarm behaviors that individual drones cannot achieve alone. The biometric identification component showed that pilot identity can be inferred from control patterns, opening possibilities for personalized drone interfaces and security applications.

Future Directions: The foundation laid by this project continues at TARS, with ongoing work on automatic assembly using manipulator robots, parametric additive manufacturing for replacement parts, and advanced swarm coordination algorithms. The vision is to create fully autonomous drone swarms that can self-assemble, self-repair, and adapt to changing mission requirements while keeping a human in the loop through VR interfaces.

Georgia State University Immersive Computing & IoT REU: June - July 2021

Undergraduate Researcher

Developed fast 360-degree video compression based on user attention, using ML for IoT sensing. Learned technologies related to immersive computing, data science, & game programming under Dr. Zhisheng Yan of GSU.

Project: Advanced 360 Video Compression via User Attention Metrics for IoT Camera Sensing
Institution: Georgia State University
Mentor: Dr. Zhisheng Yan
Duration: June - July 2021
Role: Lead Undergraduate Researcher
Project Details: Explore GSU REU Researchers
Competitive Edge: Selected from among the top 3.42% (10 out of 292)

Introduction: I likely got selected because of the interesting work I was investigating at TARS on VR Drone Teleoperation. I always encourage people to do things that are out of the ordinary, it's one way to stick out when you need to, all I do is follow my curiosity and it takes me pretty far. Out of 292 applicants, I, along with 9 others, was accepted into a very unique and fascinating undergraduate research program with a 3.42% acceptance rate that allowed me to learn about GPU programming, immersive computing (virtual reality ofc.), the "internet of things", sophisticated computer vision techniques for image compression and image representation and feature extraction. Along with this REU, I was simultaneously taking an REU on Smart and Autonomous Internet-of-Things (IoT) Systems. You can probably tell I like doing two things at the same time… well, in this case, the professor I was initially supposed to work with wasn't responding to my emails, so they had me on both REU's simultaneously, worked out for me I suppose!

Since this was the pandemic, I was virtually interfacing or zooming as a zoomer is wont to do with people- during the Georgia State University REU. So basically, I got to look outside the matrix. I decided to know the truth about our digital world, its infinite variety of images, so I took the red pill of image understanding, ah yes, the truth! Images were just matrices, after all.

Ah yes, there's nothing I would rather do that summer than learn about image compression. I remember the birds chirping, what do they sing? Oh they tell of course of the fascinating world of 360-degree video compression. Basically, when you look at an image, or a video, you only look at some things, other things don't really make our brain that curious, they fail to capture our attention. Our algorithmic society often tries to harvest our attention through recommender systems. It's true! I read it in a book somewhere, maybe, maybe not. The shareholders of random company finds it most important that random citizen # 481,283 keep streaming Rick and Morty on Netflix, or… else!

You can actually create models of this attention or saliency process, and you can map them onto videos, which are just sequences of images. My mentor Zhisheng Yan did just that, and created saliency map datasets here. This allowed us to create a machine learning model to predict which chunks of an image should and shouldn't be compressed. Why waste resources on data that nobody cares about? Besides, saving bandwidth is a huge deal for energy strapped 360 degree video sensors (like the cameras that capture footage of wildlife 24/7).

You can find the project website here

Project Concept:
Mentor: Dr. Zhisheng Yan (zyan@gsu.edu)
Title: 360 Video User Attention Detection through Spatial Audio
Description: 360 videos, namely VR videos, are large in terms of data amount. They have multiple viewports while a user only views one viewport at a time. There is a lot of redundant data that needs to be computed and transported in 360 videos. Identifying user attention in 360 videos has been an important research topic. By knowing which viewport user will be looking at, we can develop more efficient foveated rendering or streaming systems to improve user experience while saving computing and networking resources. In addition, understanding user attention can produce better 360 movies by focusing on content that users are more interested in.

Head movement data and video content features have been used to identify user attention, namely visual saliency, in 360 videos. In this project, we focus on a completely new perspective by leveraging spatial audio data to infer visual saliency. We will collect the user's head movement data when viewing 360 videos that have spatial audio. We will utilize machine learning models to identify the relationship between the head movement and the spatial audio.

References:
A spatial audio dataset
360 video user attention detection through video content features
Code for Saliency Maps
Our code for blazing fast 360 degree video compression

Out of 292 applicants, Jakir was one of only 10 people selected for the program this summer.
"I chose Clarkson University because I understood that it would be a space where I could have access to the opportunities that would allow me to explore my potential as a budding scientist. The most important thing that attracted me here was how easy it was for undergraduates to get involved with research," Jakir said. Read more about it here.

Education

Clarkson University

B.S. in Computer Science | B.S. in Business Intelligence & Data Analytics | Minors in Robotics, Math

May 2023 | Potsdam, NY

Relevant Classes

A selection of classes I've found particularly interesting and impactful:

Database SystemsOperating SystemsAlgorithms and Data StructuresDeep LearningAdvanced Topics in Deep LearningComputer Graphics3D Digital ModelingComputer VisionHuman-Computer InteractionDatabase Design & ManagementCalculus IIIDifferential EquationsApplied Linear AlgebraDiscrete Mathematics & ProofProbability and StatisticsRobotics I, IIBig Data ArchitectureComputer ArchitectureApplied Data AnalyticsData Driven Analysis of Complex SystemsIntro to Complex NetworksStrategic Human Resource ManagementOrganizational BehaviorNeural Circuits for CognitionEmergent Computations Within Distributed Neural CircuitsTinyML and Efficient Deep Learning Computing

Skills

C/C++PythonPyTorchGPT-4WebGLJavaC#JavaScriptRMatlabSQLFlaskMayaExcelTableauScikit-learnML-AgentsMatplotlibVirtualBoxBlenderAdobe SuiteGit/VCSROS PythonOmniverseSolidworksUnity3DGCPAWSBCIVRArduinoTCP/IPUnixWindowsMacNLPNLTKOpenCV

Projects

XRAgents

In development open source framework for high fidelity 3D Agents in XR worlds.^[11]

Open-source framework in Unity/C# with OpenAI API for 3D AI agents that live in XR worlds: spatial reasoning, natural language, persistent memory across sessions. Sub-100ms response times for agent actions. The first real application was neurosurgical training: AI avatars on HoloLens 2 with Kinect capture and WebSocket holoportation, demonstrated at the MIT Museum. Real-time body tracking and avatar sync keep latency under 100ms.

Moodspace

MIT Reality Hack 2022 | Emotionally aware meta-conversations with Hololens and EEG. Real-time biometric data-driven XR.^[12]

Team of 4, built in 48 hours. We piped an OpenBCI EEG headset (8 channels, 256Hz) into HoloLens 2 via Unity MRTK to classify emotional states in under 50ms and show them as shared context during conversations.

LifeInBetweenXR

MIT Reality Hack 2023 (semifinalist) | Life in Between: hyper-local MR, 3D building meshes, transit-aware search (Unity, ARCore, Esri ArcGIS).^[13]

Unity + ARCore + Esri ArcGIS 3D mapping; Google Maps–backed search filtered along subway/bus lines; bookmarks, MapExplore along the route, Dolby social stream. Vive XR Elite passthrough MR in the submission. Semifinalist out of 200+ teams. Full write-up: /portfolio/realityhack/2023.

EmpathyEase

MIT Reality Hack 2024 | AI in mixed reality, mediating conflicts between people, emphasizing nonviolent communication.^[14]

Quest 3 with Meta Presence Platform. We used GPT-4 for real-time conversation analysis, implementing Rosenberg's Nonviolent Communication framework as the AI's guiding structure. Spatial audio positioned each speaker in the shared MR space. 48-hour hackathon.

Society of Scientists

AGI House × Sundai.club AI Agents Hackathon (MIT), 2024: 1st place. Nine specialized agents on AI21 Jamba + Exa for grant ideation; inspired by SciAgents-style discovery research.^[21]

Team: Igor Sadalski, Elior Benarous, Alif Jakir. AG2 (AutoGen) + AI21 Jamba, Exa (optional; cached paper summaries), multi-agent grant-proposal generation extending SciAgents / SciAgentsDiscovery lineage. Maintained code: github.com/Caerii/SocietyofScientists. Judges: Tuan Ho, Mark Weber, Yaniv Markovski.

AI YouTube Philosophy Channel for Halcyox

Halcyox experiment: Python pipeline (GPT-4 scripts, ElevenLabs, DALL-E, FFmpeg) that shipped 200+ videos: 6.9M impressions, 940K views, ~$3K ad revenue on ~$500 API spend. A pipeline stress-test, not the main research line.^[15]

I wrote a Python pipeline that chained GPT-4 (scripts), ElevenLabs (TTS), DALL-E 3 (visuals), and FFmpeg (compositing) to generate 200+ videos over the course of 2023. The channel hit a 6.7% click-through rate and 3:06 average view duration, about $3K in ad revenue on roughly $500 in API costs.

VisuaML

Web-based PyTorch FX graph editor with Yjs multiplayer sync (<100ms), shape-safe pipeline export, and a Rust path to Open-Hypergraph macros: collaborative model design without everyone editing raw notebooks.^[16]

React + TypeScript frontend with Yjs CRDT for real-time sync: 50+ concurrent users, under 100ms sync latency. PyTorch FX extracts computation graphs; a Rust backend compiles them into categorical algebra for formal composition proofs. The graph editor compiles directly to executable training pipelines with shape-safe composition via graph macros. Inspired by Milewski's "Category Theory for Programmers" and Shannon's communication theory.

CollaborationCircuits

Pre-registered mechanistic interpretability on social cognition and multi-agent circuits in LLMs (theory of mind, interference, cross-model validation, interventions). Round 1 exploratory work on Qwen3-4B is archived under archive/round1_explorations.^[17]

Round 2: EXPERIMENTAL_PLAN.md at repo root. Round 1: ~20 exploratory experiments (ToM, belief tracking, circuit discovery) plus MATS writeups in archive/round1_explorations: github.com/Caerii/CollaborationCircuitsMechInterp/tree/main/archive/round1_explorations. circuit-tracer: I maintain a production fork (PyPI, docs, CI) of Decode Research's research library; see alifjakir.com AI Civilizations → Measurement → Collaboration circuits. Related: attribution graphs (Ameisen et al., Lindsey et al., 2025).

OpenGalea

MIT Reality Hack 2025 | OpenBCI + Quest 3 neuroadaptive MR, colocated multiplayer. Meta prize (five Quest 3s). Full write-up on site.^[18]

Team of 5, 48 hours. Quest 3 with OpenBCI Galea (EEG + fNIRS + EMG), Unity Netcode for local multiplayer. We got 85% accuracy on binary thought classification (left/right) and built shared AR experiences controlled by brain signals. The hardware is 3D-printable, the software is open-source, and the whole system costs 15.8x less than commercial BCI rigs. Finalist prize winner. Supports single-player and multiplayer via Meta Colocation APIs.

Personoid (The Future is Chrome)

MIT Reality Hack 2026 | De-Escalate: Expressive Humanoid Persona. GalaxyXR → Booster K1, LLM keyframes, Gemini Live. Google DeepMind + Reality Amplifier. Full write-up on site; Devpost has gallery + submission.^[20]

Team of five. Unity 6000 Android XR on Samsung GalaxyXR, Booster Robotics SDK (DDS), FastAPI keyframe server, LLM providers (Gemini, OpenAI, Anthropic), React + Three.js 3D preview, Gemini Live integration. I led Booster integration: Python keyframe system extending the SDK, Unity C# service layer, kinematic conversion, server wiring, documentation, and voice-to-robot control paths.

NASA RASC-AL Competition

Space Mission Concepts | Led large international collaboration (Sep 2021 - May 2022) (HKUST, Khalifa, RMIT, Kyushu) for scalable safe autonomous self-charging nuclear modular storage infrastructure system for Lunar South Pole missions. Adhered to NASA Systems Engineering principles.^[19]

I led a team of >50 people across 4 universities (HKUST, Khalifa, RMIT, Kyushu) over 8 months, coordinating technical design reviews and system integration across time zones. We designed a 10MW Kilopower reactor system (45 tons, 220 W/kg, 15-year lifespan) for Lunar South Pole power infrastructure, following NASA Systems Engineering Handbook (SP-2016-6105).