Measure before mythologising.
Every claim needs a baseline, a protocol, and evidence that can survive contact with production.
Independent AI R&DEst. 2026 / Open systems
We build what
doesn't exist yet.
An independent research and engineering group working on agent systems, model adaptation, evaluation, and production AI.
Explore the research
Founder study 001 — research / systems / production
Agent systems ✦Model adaptation ✦Evaluation ✦Context engineering ✦Voice ✦Infrastructure ✦Agent systems ✦Model adaptation ✦Evaluation ✦Context engineering ✦Voice ✦Infrastructure ✦
00 / Position
Most “AI R&D” is expensive API integration. Useful, sometimes. Research, rarely.
We enter where the implementation path is not obvious: when the system needs a new architecture, an honest experiment, a model intervention, or evidence strong enough to stop a bad direction.
The output is not a deck. It is a working system, benchmark, dataset, protocol, decision, or reusable capability.
01 / Selected research
Systems with a thesis.
Open implementations, explicit constraints, reproducible evidence.
01
Agent engineering control plane
Supervisor
A local system that turns coding agents into an accountable engineering team: persistent product context, role-owned lifecycle gates, isolated worktrees, review, verification, and controlled integration.
Codex, Claude Code, and Qwen adapters · production and end-to-end gates · Apache-2.0↗
02Coordination infrastructure for coding agents
Agent Bus
A production-oriented Kotlin message bus where local and remote coding agents discover peers, coordinate through topics, exchange direct handoffs, and keep bounded offline queues across interrupted sessions.
Ktor HTTP + SSE broker · MCP adapter · disk persistence · JUnit 5 · GitHub Actions↗
03Experimental knowledge for coding agents
Agent Experiment OS
An MCP-native system for turning agent runs into reusable knowledge: hypotheses, test design, observed failures, metric movement, interventions, and accumulated policy.
Postgres + pgvector · matrix experiments · source-backed work briefs · public research preview↗
04Mixture-of-experts model research
MoE Surgeon
A fail-fast toolchain for scanning, profiling, pruning, evaluating, and exporting MoE checkpoints without mutating the source model tree or hiding topology mismatches.
Gemma 4 26B-A4B · GGUF · Qwen3.5-MoE · deterministic artifacts and checksums↗
02 / Research index
01Agent Eval HarnessTrajectory-level evaluation↗02Injection GuardPrompt-injection defense↗03Context TelemetryContext burn-rate analysis↗04Context CompactorMeasured tool-output compression↗05MCP Tool RetrieverTop-K tool-schema retrieval↗06Meta HarnessHarness improvement from traces↗07Trajectory TipsReusable lessons from agent runs↗08ERL MemoryExperiential reflective learning↗09LLM on RailsAgent-native Rails engineering↗10JootjOOQ test-data factory↗
03 / Product work
Product thinking, shipped.
We do not just ship software. We understand the product behind it — from user decision and business logic to architecture and production.
03A / R&D Club products
Products we build to sell.
Commercial systems conceived, researched, and built inside R&D Club — where our own product thesis carries the risk.
03B / Client products
Product systems built for others.
End-to-end products created for external organizations: business model, user decisions, system architecture, and production delivery.
АГКА Отчётность
A configurable reporting system for a distributed professional organization.
View case ↗
Randevu
One operating system for bookings, clients, schedules, and a business website.
View case ↗
Uppora
A donation platform designed around the supporter, without mandatory subscriptions.
View case ↗
Jubitom
Digital commerce for a complex luxury catalog connected to a physical retail network.
View case ↗
MemBet
On-chain gaming on Solana with wallet-native identity and realtime social UX.
View case ↗04 / Operating principles
Constraints are part of the model.
Latency, noise, budget, operators, and failure recovery belong in research from day one.
Own the whole product.
User value, business logic, UX, architecture, and production behavior are one engineering problem.
02A / Journal
All notes ↗Notes from inside the work.
Positions, experiments, engineering decisions, and conclusions we can defend.
05 / Contact