scroll down to know about my project
This proflie is a mini research and tools lab for AI resrchers
I'm Kapil Dhinwa, a school student learning AI by building practical projects with AI-assisted development.
I enjoy exploring new ideas in machine learning, language models, memory systems, neuromorphic computing, and wave-based AI. I design the ideas, break down the problems, and use AI as a development partner to help turn them into working projects.
- fluid-ARC β A numerics-first language model built on sheaf-guided spectral diffusion β every math claim machine-checked (21 self-checks, 59 CI tests).
- DNA_DYNIMO β An experimental PyTorch toolkit exploring Fourier-feature neural fields for matrix generation, budget-aware routing, and a compact causal transformer.
- NEXUS-16 β A wave-native, tokenizer-free byte-level semantic encoder crunching text at ~956K tokens/s.
- Spiking_GNN β A bio-inspired 241M-parameter spiking graph neural network that computes with spikes, not dense activations.
- FREQUENCY_EVER β A wave-native blueprint: no floats, no tokenizer, no GEMM β computation is wave interference.
- fluidtech π flagship β A bounded memory layer for language models with exact recall, bounded KV-cache management, and an optional PyTorch adapter.
- TMU β A GPT-2 plugin layer that carries world time into a frozen base model β only the layer trains.
An honest archive of an ambitious attempt to compress full LLM knowledge and behavior into <1β10 MB β documenting what worked in narrow tests, where the approach broke, and why I pivoted.
- DNA-LLM (Part 1 β Weight Generation /
DNA Nanna): Explored Implicit Neural Representations (randomized Fourier Features + MLP) to map 2D coordinates into ~1B virtual parameters (3.75 GB) from ~257K DNA weights (~1 MB, 3913x ratio). Why it fell short: Generating structured weights from coordinates didn't translate into compressing real pretrained LLM knowledge without quality collapse. - DNA-LLM-PART-2 (Part 2 β Behavior Extraction /
Reverse.Dean): Shifted from static weights to activation-level "Behavior DNA" (8,784 LOC across 21 experiments using runtime SVD bottlenecks, logit lens, concept injection, spiral zoom, and distribution-shift tests). Why it fell short: Activation bottlenecks and concept rescue couldn't generalize reliably to unseen, open-ended prompts. - DNA-LLM-PART-3 (Part 3 β Knowledge Memory /
HELIX): Separated storage from interpretation using SPIRAL content-addressing, wave lattices, Reed-Solomon ECC, and relational rule algebra (2,308 LOC). Result: Achieved 99.9β100% accuracy in 3β13 KB on narrow, rule-dense domains (e.g., structured geography/chemistry relations), but failed the main goal β it does not scale to general LLM knowledge, full reasoning behavior, or billions of open-ended facts.
Key Takeaway:
- Rule-dense, structured domains compress dramatically well, but general natural-language intelligence and reasoning cannot be replaced by compact coordinate/hash/rule memory alone β which motivated my shift toward practical plug-and-play memory layers like
fluidtech.
And all the toolkit(24 repos) and advance tools(94 repos) made by above mentioned 3 failed projects ...
Small, honest, CI-verified research tools β every repo ships with bounded tests, self-diagnosing CI, CodeQL, and honest limitations sections. Negative results included.
Measurement & trust:
- WAVE-CLASH β Stack many messages into one wave signal and measure the crosstalk when you read them back.
- NEVER_LIE β A digest-guarded exact memory that abstains on corruption instead of answering wrong.
- MOLECULE-COMPRESS β Quality-gated LLM compression experiments; first result: ternary PTQ collapse, measured on Qwen2.5-0.5B.
- VARIEANCE-LIES β Rank sweeps showing how much of "99% variance preserved" is a lie for actual behavior.
- RULE-FORAGER β Mines inverse relations between facts, so derived rules cost zero stored bytes.
- POWERLAW-PROBE β SVD spectra of activations vs weights β tells you if a layer is compressible at all.
- LLM-EQUATION β A curve distiller that fits equations with an extrapolation gate β and rejects curve-fits that lie.
- ANT-HILL β Builds a 3D "occupancy mold" of how a model physically organizes knowledge, with an HTML viewer.
- Draft_ponit_harness β drift-point: the exact generated token where a compressed model replay first diverges from the original.
Model introspection:
- TRACE-PATH β Finds which block of a model carries a piece of information, with activation patching.
- BUS-TRACE β Shows who writes what into the residual stream, block by block.
- HEAD-DETECTIVE β Classifies what every attention head actually does, straight from its attention patterns.
- LOGIT-STREAM β Gives every answer token a birth certificate: which layer pushed it to the top.
- STATE-MIRROR β Watches where two nearly identical prompts start to differ inside a model.
- -KNOCKOUT-LAB β Removes one block at a time and ranks every component by the damage it prevents.
- NOISE-STORM β Pressure-tests every block with calibrated noise and maps what is fragile.
- PROBE-STACK β Maps where a property lives inside a model, one tiny probe per layer.
- INNER-VOTE- β Checks whether a model agrees with itself: hidden-state consensus across paraphrases.
- KNOW-VAULT β Censuses the storage vault inside a memory layer: how many slots, how many dead.
Weight & representation audits:
- LAYER-CENSUS- β Per-layer atlas of weight matrices: shapes, effective rank, spectrum decay, norm statistics.
- TWIN-TOWER β Detects duplicated and redundant layers with cross-layer similarity clustering.
- DEAD-CELL β Maps internal waste: zero rows and columns in weights, plus neurons that never activate.
- SKIP-RUN- β Measures how much model depth is real by skipping every k-th block and tracking behavior deltas.
- GEOMETRY-SCAN β Per-layer hidden state geometry: intrinsic dimension, anisotropy cone strength, variance spectrum.
β¦ These projects are co-built with AI, but the ideas, direction, and system designs are mine. I use AI to help execute, test, and iterate on the concepts I want to explore.
All repositories feature: CI workflows, CodeQL security analysis, comprehensive tests, professional documentation, and bugless Python code.
DNA & Fourier-based Systems:
- FourierNet β Fourier-based weight generation with extreme compression (3913x)
- DNA-Encoder β Encode any model into DNA representation
- DNA-Pro β Production-ready DNA generator with inference mode
- DNA-Lite β Lightweight DNA for small models
- SparseDNA β Sparse matrix generation with DNA
- QuantDNA β Quantized DNA weight generation (1-8 bit)
- MatrixGen β Generate matrices from coordinates
- WeightForge β Forge weights from minimal parameters
- ParamShrink β Parameter shrinking via Fourier features
Compression & Activation Analysis:
- CompressorNet β Neural network compression benchmark suite
- ActivationCompress β SVD, PCA, and random projection compression
- BehaviorDNA β Behavior analysis and compression
- SVD-Lens β SVD-based layer analysis and bottleneck detection
- BottleneckTest β Bottleneck testing framework
- DistributionShift β Distribution shift detection
- GeneralizationTest β Generalization testing
- OutputSensitivity β Output sensitivity analysis
- ActivationRank β Activation rank analysis
- SensitivityMap β Sensitivity mapping
- VarianceBasis β Variance basis analysis
- KnowledgeProbe β Knowledge probing
- GradientFlow β Gradient flow analysis
Knowledge & Memory Systems:
- KnowledgeGraph β Knowledge graph construction
- InterferenceEngine β Interference detection and mitigation
- ChainReasoner β Chain-of-thought reasoning
- AliasDeriver β Alias derivation
- FactBase β Fact-based knowledge storage
- RuleEngine β Rule-based reasoning
- ContentAddress β Content-addressable memory
- ECC-Memory β Error-correcting code memory
- CausalMemory β Causal memory system
- SemanticMemory β Semantic memory system
- TemporalMemory β Temporal memory system
- HierarchicalMemory β Hierarchical memory system
HELIX Architecture:
- HELIX-Core β HELIX core architecture
- SpiralDB β Spiral database for HELIX
- WaveMemory β Wave-based memory
- HELIX-Pro β Production-ready HELIX
- Fourier-HELIX β Fourier-based HELIX
- Compression-HELIX β Compression-optimized HELIX
- DNA-HELIX β DNA-based HELIX
- Activation-HELIX β Activation-based HELIX
- Behavior-HELIX β Behavior-based HELIX
Router Systems:
- ComplexityRouter β Complexity-based routing
- KnowledgeRouter β Knowledge-based routing
- MemoryRouter β Memory-based routing
- SpeedRouter β Speed-optimized routing
- HybridRouter β Hybrid routing
Benchmark Suites:
- CompressionBench β Compression benchmarks
- KnowledgeBench β Knowledge benchmarks
- MemoryBench β Memory benchmarks
- SpeedBench β Speed benchmarks
- RobustnessBench β Robustness benchmarks
Serving Systems:
- DNA-Serve β DNA serving system
- HELIX-Serve β HELIX serving system
- Compression-Serve β Compression serving
- Knowledge-Serve β Knowledge serving
- Router-Serve β Router serving
Libraries:
- DNA-Lib β DNA library
- HELIX-Lib β HELIX library
- Compression-Lib β Compression library
- Activation-Lib β Activation library
- Behavior-Lib β Behavior library
Research Tools:
- DNA-Research β DNA research tools
- HELIX-Research β HELIX research tools
- Behavior-Research β Behavior research tools
- Activation-Research β Activation research tools
- Memory-Research β Memory research tools
Testing Frameworks:
- CompressionTest β Compression testing
- MemoryTest β Memory testing
- SpeedTest β Speed testing
- RobustnessTest β Robustness testing
Generation Systems:
- InfiniteGen β Infinite generation
- ChunkGen β Chunked generation
- SafeGen β Safe generation
- StreamGen β Streaming generation
- ParallelGen β Parallel generation
DNA Architectures:
- DNA-Transformer β DNA-based Transformer
- DNA-MLP β DNA-based MLP
- DNA-CNN β DNA-based CNN
- DNA-RNN β DNA-based RNN
- DNA-GAN β DNA-based GAN
Additional Tools:
- NeuralCompress β Neural compression tools
- DeepAnalyze β Deep analysis tools
- ModelLens β Model introspection lens
- KnowledgeNet β Knowledge network
- MemoryFlow β Memory flow analysis
- ActivationNet β Activation network
- BehaviorNet β Behavior network
- HELIX-Net β HELIX network
- DNA-Net β DNA network
- CompressNet β Compression network
- RouterNet β Router network
- ServeNet β Serving network
- Python and PyTorch
- Large language models and memory systems
- Neural networks and graph learning
- Neuromorphic and wave-based computing
- Reproducible open-source experimentation
- π School student
- π§βπ Learning: AI and machine learning
- ποΈ Building: projects that excite me
- π± Back online: Holi 2027
I enjoy experimenting with AI architectures and learning in public. If you find an issue or have an idea, feel free to open an issue in the relevant repository.
Learn. Build. Share. Repeat.

