A
Can a compact state preserve what matters?
EILER explores how evolving state and selective memory updates might work together in sequence models. The value of that design has to be measured against strong baselines.
INDEPENDENT AI RESEARCH & ENGINEERING
Building systems
that hold together.
Ideactivate is a founder-operated AI research and engineering lab. We build developer infrastructure for AI-assisted software work and investigate how systems retain useful state, expose structure, and verify outcomes.
LAB NOTE / 001STATE → MEMORY → OUTPUT
FIG. 01RESEARCH MOTIF / NOT A BENCHMARK
AI SYSTEMS✳MEMORY & STATE✳DEVELOPER TOOLS✳EXPERIMENTAL SOFTWARE✳AI SYSTEMS✳MEMORY & STATE✳DEVELOPER TOOLS✳EXPERIMENTAL SOFTWARE✳
01 / RESEARCH AGENDAQUESTIONS FIRST
Research starts with
a system problem.
We turn technical questions into working systems, then make the assumptions and limits visible.
The current work spans structured context for coding agents, recurrent sequence-model research, and the architecture of complex creative software. These projects are at different stages; each page distinguishes implementation from hypothesis and future work.
↗
B
What context should a coding agent see?
Atrium explores repository structure, local memory, and verification tools for AI coding workflows. The goal is better task outcomes—not a smaller prompt at any cost.
C
How do complex tools stay predictable?
FluxCut explores a modular foundation for a timeline-based media editor, where commands, project state, and rendering can evolve without becoming tangled.
02 / PROJECTSTHREE LINES OF INQUIRY
From question to
working code.
The projects are independent, but share one standard: explain the mechanism, test what can be tested, and let evidence—not presentation—carry the claim.
DEVELOPER INFRASTRUCTURE · RUST · OPEN SOURCE
Atrium
Give coding agents structured repository context and a path to engineering feedback.
AST graph · Local memory · Verification↗
MODEL ARCHITECTURE · EXPERIMENTAL RESEARCH
EILER
Test how evolving state and selective memory might work together in sequence models.
Complex SSM · Delta memory · Recurrence↗
NATIVE SOFTWARE · RUST · TAURI · WGPU
FluxCut
Build a modular media workspace with explicit timeline state, command flow, and a rendering path.
Timeline · Command bus · Rendering↗
✳ Projects are developed independently and may be at different stages. Descriptions reflect design intent and current exploration, not independent validation of performance.
03 / FLAGSHIP BUILDAI-ASSISTED SOFTWARE ENGINEERING
Tools around the model
matter, too.
Atrium is the most immediately practical expression of Ideactivate’s work on AI systems: give a coding agent structured repository context and a route to observable engineering feedback.
Through an MCP-facing tool interface, Atrium is intended to work with Claude Code and other compatible coding clients. The model remains responsible for reasoning; Atrium supplies repository-specific tools and local checks. The open question is whether this structure improves real task outcomes once overhead and failure cases are counted.
01
Repository structure
Expose symbols and relationships that help a coding client navigate a repository without treating every file as an undifferentiated text block.
02
Context with provenance
Keep selected project facts local and inspectable, with the repository—not cached notes—remaining the source of truth.
03
Feedback that can be checked
Return local build or test outcomes to the workflow, while being explicit that passing checks are evidence—not a guarantee of correctness.
04 / THE LABSMALL BY DESIGN
INDEPENDENT
RESEARCH &
ENGINEERING✳IDEACTIVATE / LAB
RESEARCH &
ENGINEERING✳IDEACTIVATE / LAB
Small by design.
Serious about evidence.
Ideactivate is founded and operated by Shreshth Banjare. It is an independent, founder-operated lab—not a claim of a large team, institutional funding, or completed frontier-scale research.
The work is built in public where possible. Implemented behavior, research hypotheses, known limitations, and planned milestones are kept distinct so a reader can judge the work for themselves.
The aim is straightforward: make difficult ideas concrete, create useful engineering artifacts, and improve them through evaluation and feedback.
Find the work on GitHub ↗OPEN TO GOOD QUESTIONS
Have a system problem worth exploring?
Open-source issues, research questions, and engineering collaborations are welcome. A clear problem statement is the best place to start.
shreshthbanjare@ideactivate.cloud ↗