JARVIS
A wearable AI, built as one system.
Independent builder
I build AI
you can wear.
Wearable systems, EEG models and spatial intelligence. From the first circuit to a working agent.
Design visualization. Actual prototype photographs follow below.
What I've built.
Hardware you can see. Research you can inspect.
Explore the system
HarmonicFoldNet-SSVEP
A compact
EEG Transformer.
+6.75 ppBETA, 1.2 s vs. protocol-adapted SSVEPformer
HarmonicFoldNetSSVEPformer
Public-data study. Selection-aware comparison.
Spatial computingMemory needs
Memory needs
a place.
1,281recorded keyframesWatch the field demo
One direction.
Three ways in.
I build systems that connect perception, memory and action. JARVIS is the working foundation; EEG and spatial computing extend how people can interact with it.
JARVIS. An AI that shares your world.
A glasses camera, local phone inference and a memory-based agent form one working system.
Perception becomes context.
Time-aligned vision, speech, location and message summaries feed personal memory. Habit-triggered semantic review decides when an action or reminder is useful.
Snapdragon 8 Gen 3
- ~600ms
- Fastest single-frame result
- ~1s
- Average single-frame inference
NVIDIA RTX 5060
- ~300ms
- Fastest single-frame result
- ~600ms
- Average single-frame inference
Running measurements supplied by the builder, October 2026. Not independently re-benchmarked for this portfolio. Single-frame inference only, excluding transport, retrieval and decisions.
Measurement scopeOne shared context.
Specialized agents.
Explore the path from a camera frame to a contextual action.
GlassesCamera & audio
Local perceptionQNN / HTP
Shared memoryRetrieve & summarize
Reviewed actionEvidence & ACK
Architecture walkthrough, not a live device connection.
Glasses cameraPhone VLMTemporal evidence
The phone runs local visual inference with QNN/HTP. Frame IDs align descriptions with visual-token identity results. Temporal checks retain evidence across observations.
SummarizeHybrid retrievalShared context
Summary and retrieval agents reduce context load. Local bilingual embeddings, lexical retrieval, reranking and time/source metadata support contextual recall and object search.
Habit stateSemantic reviewVerified action
Generated habit state machines provide a fast trigger. The main agent reviews the complete observation before intervening. Tool completion relies on execution evidence and device acknowledgements.
Implementation and contribution
Camera/audio transport, glasses UI, Android local VLM integration, temporal visual reasoning, memory retrieval, real-time speech context and tool execution are implemented. Foundation models and speech services are dependencies. The RV1106B prototype uses a development kit; custom hardware development is ongoing.
Visual-token matching reuses features already produced by the local VLM. It supports learned places, negative examples, region records and same-frame identity supplementation. Appearance matching and object ownership still need scenario-specific validation.
HarmonicFoldNet-SSVEP
Foldable convolution, harmonic-biased attention and a compact Transformer for cross-subject EEG decoding.
247unique public-data participants
435kfolded executable parameters
3.04msmedian single-thread CPU inference at 1.2 s
49,000folding checks, zero label changes
Evidence you
can explore.
Compare observation windows and public datasets. Each neural result averages participant-disjoint evaluation over three training seeds.
1.2 s68.70%HarmonicFoldNet balanced accuracy+6.75 ppversus protocol-adapted SSVEPformer
HarmonicFoldNetSSVEPformer
Local branches merge.
Attention stays.
Explore how the local mixer changes at deployment. The reduced-token Transformer and harmonic attention remain in the model.
EEG
ConvIdentityNormΣ
Merged local operatorReduced tokensHarmonic attention
Shared
scorer
scorer
435,043deployment parameters
49,000 / 0folding checks / label changes
Schematic, not an attention heatmap. Equivalence checks cover the retained checkpoints and tested windows.
Benchmark35 participants
BETA70 participants
Wearable102 participants / dry & wet
Kim2025BetaRange40 participants
Architecture, adaptation and module contributions
Foldable local temporal mixing precedes time-domain and complex-spectral candidate evidence. Reduced spectral tokens feed harmonic-biased late attention and Transformer candidate modeling. An optional adapter updates just 113 parameters for participant calibration.
In the registered BETA ablation, including local temporal mixing improves accuracy by 3.99-7.23 percentage points across the tested windows. Including late attention lowers it by 0.66 points at 0.4 s and improves it by 1.20-2.16 points at 0.6-1.5 s.
Evaluation boundaries and reproducible values
All four datasets are public. Benchmark and BETA are selection-aware because later candidate screening consulted them. Wearable dry/wet analysis was frozen after model retention. The model is not the best at 0.4 s; ordinary SSVEPformer also runs faster on the measured laptop.
CPU times use batch one, float32 and one inference thread, with 100 warm-up and 1,000 measured runs. They include registered model preprocessing, exclude signal acquisition, and do not establish online or glasses latency. No final glasses-mounted electrode montage has been validated.
| Window | HarmonicFoldNet | SSVEPformer |
|---|---|---|
| 0.4 s | 27.37% | 30.26% |
| 0.6 s | 43.47% | 43.46% |
| 0.8 s | 55.60% | 52.79% |
| 1.0 s | 63.56% | 58.52% |
| 1.2 s | 68.70% | 61.95% |
| 1.5 s | 73.12% | 65.09% |
AI with a sense of space.
Visual recognition answers where. A persistent spatial model can also describe how places, objects and people relate.
From recognition
to spatial memory.
The scenic navigation prototype combines recorded maps, visual relocalization, route instructions and glasses display. The recorded route below contains 1,281 keyframes.
Recorded route playback, not live localization. Monocular map coordinates are not metre-scale distances.
Mingzhi SquareNingxia Luoying Pavilion
Gaussian splatting
and spatial AI.
The next research step explores scene reconstruction with Gaussian splatting, stable spatial anchors, occlusion and spatially grounded AI reasoning. The aim is to place assistance in the surrounding environment and let memory refer back to it.
Gaussian reconstruction is an active research branch. A complete validated spatial-computing integration is still ahead.
What I'm working toward.
EEG, temporal AI and spatial computing are parallel directions with distinct validation steps.
Glasses-compatible EEG
Compare posterior scalp baselines with peri-auricular and temple recordings. Test repeat wearing, drift, motion artifacts, personal calibration and no-command rejection before selecting a practical montage.
Candidate locations and first experiments
Compare distinct locations TP9 and T9, TP10 and T10, FT9 and F9, FT10 and F10. Nz is optional; O1, O2 and Oz remain posterior controls. These are candidate comparisons, not interchangeable site names or a validated glasses design.
Earlier public-data scalp/ear experiments used 12 participants and 1,440 unique trials. The 24-channel scalp study and ear study share the same participants. They form a separate montage feasibility study, not an additional 24-person cohort.
Behind-the-ears SSVEP studyMobile scalp and ear EEG datasetTemporal perception models
Train models that connect successive observations, object changes and user actions. Evaluate earlier intent recognition, coherent memory and timely intervention under compute and false-trigger budgets.
Spatially grounded assistance
Combine scene geometry, semantic objects and persistent memory. Test stable world anchors, incremental map reuse and object retrieval that can guide a user back to the last supported location.
Optional silent interaction
Evaluate deliberate silent articulation with surface EMG or tissue-conducted sound as a separate future input study. These signals require their own hardware and controlled data; existing EEG results do not establish this capability.
Models built, adapted and deployed.
Different contributions, with the implementation and evidence made explicit.
HarmonicFoldNet-SSVEP
Architecture & trainingFoldable local-to-global EEG decoder, multi-dataset evaluation, three seeds, ablations, participant adaptation and deployment equivalence.
TinyVLM / LoRA
Alignment & fine-tuningVisual-to-language connector training and LoRA experiments. Saved checkpoints and a completed ten-epoch second-stage log document the training work.
Visual encoder student
Distillation pipelineTeacher-guided visual distillation, projector alignment and generation smoke tests. The existing evidence supports a training pipeline, not a finished large-scale pretrained model.
On-device VLM & token reuse
Runtime integrationQNN/HTP integration on Snapdragon 8 Gen 3. Place and identity matching reuse local visual features, with frame alignment and unknown-rejection research.
Public work, open to inspection.
Code and model releases offer a way to examine the work beyond this page.
EEG decoding / reproducibility
HarmonicFoldNet-SSVEP
Source-available, noncommercial research release. Code, weights, protocols and derived source data.
LLM agents / scenario modeling
Social Simulation
A scenario sandbox with selectively activated agents and bounded model-call budgets. Exploratory results, not forecasts.
CUDA / compatibility engineering
SLAM on RTX 50
Build and C++/CUDA compatibility work for an existing SLAM system on sm_120 GPUs.
Computational neuroscience: MaleCNS
A connectome-constrained LIF simulation constructed from official male fly CNS data, with 166,700 annotated neurons and 25,582,938 directed connections. Stimulus, no-stimulus and sign-policy sensitivity conditions were tested.
This is a computational research branch. Synapse counts are not measured conductances, and the simulation is not yet a validated EEG decoder.
Official MaleCNS sourceA longer
practice of building.
The first experiments were small. Display optics, motion sensing and simple mechanisms led toward the present wearable system.
Reflection display
ESP32, OLED and a prism formed an early wearable display. A later version used a camera viewfinder as its display source.
Gesture lighting
ESP32 and MPU6050 connected motion sensing with light control.
Motorized belt mechanism
An enclosure, clamp and torque motor tested controlled belt contraction.
Devices, models and systems
Hardware development, local visual inference, memory agents and neural-interface research now share one engineering direction.