Independent work across software, machine learning, hardware, and human-computer interaction, focused on building computational systems that address practical problems and explore new forms of interaction, perception, and expression.

The projects span applied tools shaped by real-world constraints and experimental systems that investigate emerging technologies. Each moves from problem definition and technical exploration to implementation and testing, combining software development, system design, and hands-on experimentation.

The work sits at the intersection of engineering and research: building functional systems while investigating how computational methods, sensing technologies, and interfaces can extend human capabilities.

• AAgah · Inspired by MIT Media Lab's Dormio work on targeted dream incubationNeurotechnology

  A Sleep Device for Dream Incubation

  Agah is an open, DIY wearable that detects sleep onset (hypnagogia, the N1 transition between wake and sleep) and delivers timed voice prompts to bias the content of early dreams. It is inspired by MIT Media Lab's Dormio work on targeted dream incubation, which pairs physiological sleep-onset detection with audio cues. Agah reimplements the idea with low-cost, off-the-shelf components sourced from Alibaba instead of lab-grade instrumentation.

  Sensing and data pipeline

  The sensing layer is a Colmi R02 smart ring, accessed through a local-first toolchain on macOS:

  • Connects over Bluetooth Low Energy and pulls heart rate, HRV, stress score, SpO₂, sleep sessions and stages, battery state, and raw BLE packets.
  • Enables the ring's automatic HR, HRV, stress, and SpO₂ logging during setup.
  • Supports live HR/SpO₂ streaming plus BLE diagnostics for macOS connection issues.
  • Persists everything in a local SQLite database and exports CSV and JSONL for ML experimentation.
  • Generates an offline dashboard with trend charts, sleep-stage timelines, and interpretation notes.

  Incubation logic

  Physiological signals (HR/HRV trajectory, and motion where available) are monitored as the user falls asleep. When the pipeline infers sleep onset, it triggers spoken prompts or phrases at that moment, aiming to influence dream content without fully waking the sleeper.

  Limitations

  All privacy-sensitive data stays on-device. The ring's outputs, including stress, REM/deep sleep staging, and SpO₂, are consumer-grade estimates, useful for personal trends rather than clinical inference. The system does not measure hormones or cortisol and does not diagnose sleep or health conditions. Note also that a PPG ring lacks the EMG and electrodermal channels used in the original lab setup, so onset detection here is a proxy built on cardiac and motion signals and will be less precise.

  Images coming soon

• EEEG/EMG Foundation Challenge 2026 · EEG-to-IMG visual decodingVisual decoding

  Visual decoding from EEG with cross-subject transfer

  My entry in the EEG/EMG Foundation Challenge 2026 — NeuralBench Track 1 visual decoding from scalp EEG.

  Problem

  Decode visual experience from scalp EEG and retrieve the viewed image from a large candidate set. Evaluation uses top-5 accuracy in a frozen DINOv2-giant image embedding space (cross-stimulus retrieval: test images do not overlap with training).

  Approach (high level)

  I reproduced and extended the NeuralBench Track 1 pipeline: EEG encoders are trained to predict image identity via contrastive alignment (CLIP-style loss) between brain representations and precomputed DINOv2 stimulus embeddings—not raw pixels.

  Primary model: REVE (pretrained EEG foundation model from Hugging Face) with a linear probe on pooled features (official frozen probe / linear head baseline). Comparison baseline: EEGNet trained end-to-end on the same task.

  Data strategy

  • Xu2024Alljoined (~5 GB): cheap end-to-end validation of infra, metrics, and Modal workflow.
  • Xu2025Alljoined (Alljoined-1.6M, ~136 GB): 32-channel Emotiv, natural-image paradigm—closest public match to the hidden competition cohort; used for the main baseline target.

  Training setup

  • Compute: Modal serverless GPUs (A10/A100), persistent Modal Volume for data, caches, and experiment artifacts; training orchestrated from a Python Modal app (modal/app.py).
  • Stack: Python 3.12, NeuralBench 0.3.1, PyTorch Lightning, MNE-compatible loaders via neuralfetch/neuralset.
  • Protocol: Full runs use 3 random seeds (33, 34, 35) for model initialization; report mean ± std on test/full_retrieval/top5_acc_subject-agg (full candidate retrieval, subject-aggregated)—not batch-only validation accuracy.
  • Engineering: Dataset-specific fixes (e.g. duplicate trigger dedup for Xu2025 EDF timelines), NSD stimulus alignment for Xu2024, resumable downloads, detached long jobs, and experiment logging (EXPERIMENTS.md, trial manifest).

  Images coming soon

• NNeotulp · Defeating Entropy Hackathon, Society for Technological Advancement (SoTA), London LongevityRobotic surgery

  Next-scene prediction for robotic surgery

  Neotulp is a fine-tuned variant of NVIDIA's Cosmos Policy ALOHA Predict2 2B model, adapted to forecast the next visual scene in robotic surgery rather than general manipulation tasks.

  Starting from the public Cosmos-Policy-ALOHA-Predict2-2B checkpoint on Hugging Face, the project retrains the policy-style world model on surgical video so the system learns how instruments, tissue, and camera motion evolve from one frame to the next in the OR.

  The goal is anticipatory perception for teleoperated or semi-autonomous platforms: short-horizon scene prediction that could support timing, safety checks, and training simulations without replacing clinical judgment.

  The work grew out of the Defeating Entropy Hackathon at the Society for Technological Advancement (SoTA) during London Longevity. The challenge was submitted by Dylan McGuire, who went on to raise a $7.5M pre-seed led by Andreessen Horowitz to build a robotic surgeon.

  Team

  • Alara Dirik, PhD, at Imperial College London
  • Alina Mihalovits, DPhil, University of Oxford
  • Alex Vavelyuk, Data Scientist, Meta

  Base model (Hugging Face)

  Images coming soon

• OOrthoseg · Attempting to automatically segment and plan movement of 3D teeth meshes according to biomechanical principles3D modelling

  Orthodontic segmentation and treatment planning

  Orthoseg is an experiment in digital orthodontic planning — a tool built to close the gaps between what current technology promises and what clinical practice actually needs.

  Inspired by time spent building orthodontic software at 32Co, the project addresses persistent limitations in how teeth are segmented and how treatment is planned: imprecise boundaries, inconsistent outputs, and workflows that ask clinicians to work around the software rather than with it.

  Orthoseg aims to segment teeth with greater precision and plan orthodontic movement with the same rigour — turning scan data into clearer, more reliable treatment plans that reflect the nuance of each patient's anatomy.

  It is a continuation of a question first encountered in practice: what would orthodontic planning look like if the tool were designed from the ground up for accuracy, clarity, and trust?

  Images coming soon

• MMimic · Conversation patternsGenAI

  A Conversation Intelligence Tool

  Mimic listens to the patterns beneath conversation — tone, rhythm, intent — and surfaces what matters before it disappears.

  Built as a conversation intelligence tool, it turns unstructured dialogue into structured insight: themes, tensions, decisions, and the quiet signals people often miss in the moment.

  This project explores how technology can augment human understanding without replacing the nuance of speech — a tool for reflection, analysis, and better communication.

  Images coming soon

• TTaste Lab · Chemistry, molecular gastronomy, embodiment, expressionGenAI

  A food interface for translating sensation into recipes

  The project aims to help people reconnect with their senses and learn from the signals their own bodies produce. Appetite, craving, memory, and mood are treated as real information about what to eat, not noise to be overridden by a recipe database or a diet rule.

  It draws on chemistry, taste, sensation, and accumulated knowledge about food. A person describes what they want to taste in loose, poetic terms, and the system translates that impression into a dish grounded in how flavor and ingredients actually work. Each translation joins a shared, anonymous archive, building a collective record of how people turn feeling into flavor.

  It is for people who think with their mouth first.

  Images coming soon

• IImpossible Questions · Reason from first principles on a blank canvasGenAI

  A space for questions that resist easy answers

  Impossible Questions collects prompts and explorations where the point is not a tidy conclusion but clearer thinking — the kind of problems that stay open even after you have worked them honestly.

  Images coming soon

( Archive )