Skip to content
D-CSIL

Independent AI Research & Engineering

Building Intelligence Beyond the Chat Window.

D-CSIL is an independent AI research and engineering lab developing autonomous agents, novel model architectures, local AI infrastructure, persistent machine memory, and experimental systems for understanding intelligence itself.

Explore the LabView ResearchEnter Lab Console →
D-CSIL LAB STATUS
Research Systems
ONLINE
Active Experiments
01
Model Families
04
Agent Systems
03
Current Focus
HELIX-200M FOUNDATION RUN

D-CSIL explores how intelligent systems can reason, learn, collaborate, remember, experiment, and operate autonomously.

Part research lab. Part engineering workshop. Part AI proving ground. Questions matter more than trends — and claims never appear stronger than the evidence behind them.

Active Research

What We Are Exploring

01MEASURED

Novel Neural Architectures

Hybrid recurrent models (HELIX H-E-R blocks), selective state-space models, Hebbian plasticity with neuromodulation and sleep-style consolidation (SynapNet / HibbieFormer), and physics-inspired field layers (FieldTransformer).

ENTER →

02MEASURED

Local Intelligence

Foundation models trained end-to-end on hardware the lab owns — M3 Max unified memory, custom Metal kernels, memory-mapped corpora — plus edge TTS on Raspberry Pi.

ENTER →

03OBSERVED

Agentic Intelligence

Remy — the D-CSIL Lab Rat — a production terminal agent with a deterministic policy engine, reversible git-backed checkpoints, and durable memory. It replaced the earlier Hermes and OpenClaw agents.

ENTER →

04OBSERVED

Machine Memory

Cortex: shared durable memory for every agent in the lab — facts, decisions, and context that survive across sessions, served over REST and MCP.

ENTER →

05EXPERIMENTAL

Autonomous Research Systems

The Autonomous R&D Lab: objective → plan → agents → experiments → evidence → critique → revision, with claim gates and immutable audit logs.

ENTER →

06OBSERVED

Creative & Audio Intelligence

Conditional drum generation, 6-stem music source separation on MLX, XTTS voice cloning, and edge TTS — machine learning applied to sound.

ENTER →

07OBSERVED

Physical AI & Embedded Systems

Connecting intelligence to the physical world: custom electronics, sensor integration, embedded computers, and edge devices that let AI observe and act on real environments.

ENTER →

Featured Projects

Systems Under Construction

HELIXA family of hybrid recurrent language models — HarmonicStateLayer, ElasticAttentionGate, and ResonantExpertRouting blocks — trained entirely on Apple Silicon.Active Research
D-CSIL Foundation ModelsThe D-CSIL-1/2/3 training line plus the published SLM-10M: selective state-space and GLA models pretrained on all-local pipelines.Active Research
SynapNet / HibbieFormerBrain-inspired sequence models: O(L) recurrent cores, Hebbian plasticity phases, neuromodulation, and sleep-style consolidation — now converging on the HibbieFormer-HX program.Active Research
mlx-recurrenceFused Metal GPU kernels for linear recurrence on Apple Silicon: Mamba-2 selective scan, gated linear attention, RG-LRU, and rotational LRU — with full training support.Operational
Remy — D-CSIL Lab RatA terminal-first durable agent runtime. Deterministic policy engine, reversible git-backed edits, sub-agent hive with budget-aware model routing, and Cortex shared memory.Operational
D-CSIL Platform — Cortex & HarnessDurable shared memory (Cortex) and disciplined session orchestration (Harness) for every agent in the lab, served from a Raspberry Pi 5 named Bailey.Operational
Autonomous R&D LabA governed research loop: objective → plan → agents → experiments → evidence → critique → revision, with claim gates and immutable audit logs.Prototype
Multi-Agent SystemsCouncils, hives, and debate: multiple specialized models coordinating on one problem — from a Raspberry Pi MQTT hive to a five-model local debate framework.Active Research
Audio & Voice LabMachine learning for sound: conditional drum generation, 6-stem source separation on MLX, XTTS voice cloning, and edge TTS.Active Research
FieldTransformerA physics-inspired architecture: multi-scale field layers with flux, Schwarzian, and latent-diffusion regularizers over causal attention.Active Research
D-CSIL CANISA design for computational interpretation of canine communication from audio, video, pose, and interaction context.Concept
Hardware & Physical AI LabElectronics prototyping and sensor systems that connect intelligence to the physical world — from an RV leveler to machine-health monitoring and the lab's embedded fleet.Active Research

Evidence

An Active Laboratory,
Not a Portfolio

Every claim on this site links to an experiment record — including the failures. Scientific iteration is the product.

Open the Experiment Log →

HELIX-200M-F35B

HELIX-200M foundation run (helix-200m-foundation-35b)

ACTIVE

MLXR-HOTSWAP

mlx-recurrence v2 kernels hot-swapped into a live training run

SUCCESS

HF-TINYSTORIES-215

HibbieFormer TinyStories run: 215 plasticity cycles

SUCCESS

HF-CYCLE150-EVAL

HibbieFormer cycle-150 comprehensive behavioral evaluation

SUCCESS
LAB PULSEFEED ACTIVE
  • 02:14HELIX-200MCheckpoint saved · step 120,140 · train PPL 16.85 · val PPL 39.4
  • 03:02REMYOvernight session active · labrat.db WAL updated
  • 05:47ROUTER5 sub-agents / 4 providers verified · $0.02 budget enforced

HELIX-200M · TRAINING

step 120,140 · val PPL 39.4 · 2,900 tok/s

CONSOLE →

Current Questions

Questions We Are Asking

  1. Q01

    Do recurrent hybrids (H-E-R) beat transformers at matched scale on hardware one person can own?

  2. Q02

    Can plasticity-phase training (hebbian / scaling / pruning) survive contact with general text, not just TinyStories?

  3. Q03

    Does surprise-gated fast memory plus sleep consolidation cut forgetting by the pre-registered ≥20% gate?

  4. Q04

    How far can a $0.02 budget ceiling push useful multi-provider agent work toward local models?

  5. Q05

    Can a shared memory appliance make many agents behave like one accumulating lab?

  6. Q06

    What is the smallest model that earns a place on a leaderboard with a fully honest card?

  7. Q07

    Can machine learning infer communicative intent across species? (CANIS — still on paper.)

  8. Q08

    What does it take for AI to perceive and act on the physical world — sensors, embedded computers, and custom electronics included?

D-CSIL exists to pursue questions — not just products. Explore how they connect in the knowledge graph or follow the work on the research timeline.