The research threads on the current work page do not ship themselves. This page is what it looks like when those ideas meet deadlines, APIs, regressions, and other people's mental models. My GitHub tagline is blunt: human-machine symbiosis, and the problem I keep circling: most software still assumes a lone operator in front of a screen, while the work I care about is collaborative, embodied, and spread across time.
In practice that has meant shipping between XR, brain–computer interfaces, multi-agent systems, and modular robotics, not as a branding slide, but as stacks I have wired up, , and restarted. I am still learning. The repo index is messier than I want, and that is fine. I build in public on purpose.
FabuBlox (Nov 2024–Aug 2025)
First non-founding engineer at an MIT nanotech spinout. The product is a cloud platform for semiconductor process design—not generic “workflows” but structured fabrication flows where each step is a Blox (typed: RIE, ALD, spin coat, stepper litho, etc.), grouped into Sections, with reusable sub-flows called Modules that snap in via SUBSTACK bloxes.
The version-control challenge was specific: processes are ordered arrays of bloxes, so naive index-based diffing breaks when you reorder fab steps. I built a DeltaFormatter on top of jsondiffpatch that uses blox UUIDs for stable identity, treats bloxType changes as delete-old/insert-new (not a silent field mutation), and maps parameter diffs to domain-readable labels via per-blox-type schema knowledge—so the history says “In Spin Coat, Thickness was modified from 500 to 1000” instead of showing a raw JSON patch.
I also shipped PANDA (Process & Nanofab Design Assistant)—a RAG chatbot for MOSIS 2.0 fab service discovery. The search pipeline runs hybrid 35% keyword / 65% vector over 473 semiconductor tools across seven UC-system institutions, with Gemini doing entity extraction (institution, node size, wafer spec) and function calling to return clickable tool cards. Milvus stores 1536-dim embeddings; Cerebras normalizes markdown for the chat UI. The system searches and explains; it does not invent process steps.
I migrated the frontend from Create React App to Vite, getting >5× faster builds and unblocking the dev loop for the whole team. Other tooling: Auth0 + Permit.io authorization, TSOA-generated OpenAPI, Sentry, auto-generated cross-section SVGs via a stack-action engine that recomputes on every edit.
Superintelligent Group (ΣIG)
I run ΣIG now; engineering here is -native collaboration infrastructure. The platform stack: Supabase (90+ migrations), edge functions for the swarm API, Terraform on AWS, with a separate worker/runtime sandbox for execution. Three-layer split: World (durable graph of workspaces, projects, claims), Runtime (bounded execution with context packets), Interface (projections only—forum, kanban, VAC builder are views, not sources of truth).
The coordination layer uses runtime-first claim leases with durable projection to Postgres. Claims operate on named loci (repo file, symbol, text range, artifact, task, thread) with lifecycle states from requested through granted to expired or preempted. Presence splits ephemeral (cursor, viewport, attention) from durable (claim grants, mutations, tool calls) across BroadcastChannel and NATS transports. The artifact layer uses Y-Sweet/Yjs for document-like state and Loro for graph/canvas/hierarchy.
Before ΣIG consolidated the thread, a lot of this lived under Halcyox—distributed advanced HCI R&D that eventually ate everything else on the calendar. Related experiments: git2.org (swarm-native version control), SwarmIDE-Alpha (collaborative Monaco + Yjs editor), arxiv2.org (paper → living project bridge).
Hardware & hackathon builds
OpenGalea (MIT Reality Hack 2025, Meta track winner) — a colocated multiplayer neuro-MR system: OpenBCI Cyton 8-channel EEG, modified Ultracortex Mark IV frame (front parietal rings removed to fit Meta Quest 3), 3D-printed mounting brackets, gold cup electrodes at eight scalp positions. Software: Random Forest classifier on alpha/beta band power (trained on ~113K samples), Lab Streaming Layer → Python inference → UDP → Unity, Meta Colocation + Shared Spatial Anchors for multiplayer. ~$1,800/user for the full rig.
Cortana hologram (MIT Museum installation) — interactive Cortana on a Looking Glass holographic display: an Ultraleap depth sensor for hand tracking (both hands in FOV triggers the experience), AttentivU EEG/EOG sensing glasses (MIT Media Lab) for mental yes/no, Node.js WebSocket bridge between Unity and BCI. Learned the hard way: BCI disconnect kills the session unless you auto-restart the bridge.
TARS Lab hardware (Clarkson, 2020–2023) — two flying quadcopters with Raspberry Pi 4 + Emlid Navio2, ArduPilot over MAVLink, ROS Noetic + MAVROS, synced H.264 video + GPS CSV logging. Indoor flight in GUIDED_NOGPS mode using raw attitude thrust commands. Also the VRmonic piano-learning capture pipeline: multi-camera Azure Kinect depth/color alignment, MediaPipe hand landmarking at 21 joints × 3D, aggregated across sessions to HDF5 → CSV for Unity playback.
MosaicDrone — modular self-assembling omnidirectional drones: pogo-pin + electromagnet docking, RK3588 or NVIDIA Orin Nano compute, ROS 2 Humble. Descended from TARS Lab stacking experiments (pancake v1, big-box v2, small-box v3 with magnets). Architecture and simulation specs complete; no flying prototype yet in this repo. Architecture and simulation are done; flight is not.
Formative shipping (Clarkson & GDSC)
Undergrad at Clarkson University (2019–2023): dual degree in CS and Business Intelligence & Data Analytics, minors in robotics and math. Research at TARS Lab (XR, robotics, CV, ML). I ran the Google Developer Student Club as lead/organizer (Jun 2021–May 2023)—semester roadmaps, workshops on web/cloud/XR, speakers, and a culture where people went from curiosity to a demo they could show someone. That club is how I found MIT Reality Hack, which redirected the rest of the trajectory. DTU (Denmark, 2023) added nanofab and photonics—less JavaScript, more cleanroom.
Advising & collaboration
Engineering and research advising at Doppel Labs and Wallace. Collaboration with Augmentiv Labs on a scientific reproducibility stack using smartglasses to make lab procedures verifiable and repeatable. Affiliated researcher at MIT Media Lab, Fluid Interfaces. MIT Reality Hack every year from 2022–2026 (2026 DeepMind prize, 2025 Meta track winner, and earlier teams on the portfolio).
Open source and reproducibility
When I open-source work on GitHub, I want you to re-run, inspect, and extend it without fighting the packaging. David Donoho's ideal of frictionless reproducibility is the bar: collective intelligence through artifacts strangers can actually break, not only features shipped on a deadline.
Engineering values
Tools should respect deep work and idiosyncratic thinking. That is the same bar as partnership on the philosophy page, expressed as latency budgets, permissions, and audit trails instead of a nicer chat tone.
Domain and implementation stay coupled. At FabuBlox the version-control problem was not generic diffing. It required per-blox-type schema knowledge, move detection in ordered lists, and labels a process engineer could read. Knowing what an RIE etch step means was part of the engineering. A surprising amount of shipping is epistemology: what does the system believe is true, and how do we catch it when that belief is wrong?
AI coding tools are power tools. At FabuBlox they accelerated the team; they did not remove the obligation to understand nanofab semantics, auth boundaries, or what happens when PANDA a tool recommendation. I use models to explore faster; I do not outsource responsibility.
Where to start reading
github.com/Caerii — README first, then the curated tables, then the repo index (~270+ repositories). For narrative context: homepage, current work, portfolio, papers. If you are building at the research-to-product boundary—real-time systems, swarm tooling, human–AI collaboration—email or open an issue on the relevant repo.