LLM solutions for any need. Agents for human purposes.
Close the gap between LLMs and human concepts
Frontier AI sits behind expensive stacks and jargon. RYIUK experiments with and builds around cloud and local LLMs for different workflows — with a strong interest in LLM architecture, training, fine-tuning and inference prototyping: constant experiments in making models faster with less and cheaper. Machine learning sits beside the LLMs in every loop: feature stores, surrogates, ranking models — the same evidence bar either way.
In research we are introducing jev — kev, a packed-question decision layer for agentic workflows — starting with finance as the first hostile domain, then general tasks. The aim: a world running LLMs for human purposes — optimizing tasks, delivering value at every scale. Built as a research operation: we show the process and the artifacts, not promises.
Receipts, not promises
Every result ships with its run log and its kill list. We document what failed — the part most people hide.
ML + LLM, one bar
Models and agents score like the strategies do — honest validation, weak spots logged, nothing promoted on vibes.
jev — kev
Decision-layer research for agentic workflows: calibrated questions in the loop, finance first as the proving ground.
kev on the project →Local inference. Strong harness. Cheap pipelines.
Local inference on llama.cpp, paired with a strong agent harness — OpenCode, Hermes, Orca, ADE for orchestration, Herdr for the terminal plane. The output: efficient pipelines that cut human intervention without losing quality — resources, time and potential maximized, not traded away.
Cloud + local, open first
Open-source models, free-to-cheap providers and inference. llama.cpp on your machine; cloud when the workload asks for it. No proprietary lock-in.
Agents that keep the wheel
OpenCode and Hermes drive the work; Orca and ADE handle orchestration; Herdr keeps terminals and panes under one roof. Typed tools, not vibes.
Average consumer gear
Runs under average consumer hardware — no special capacity required. Efficient by constraint: less silicon, less burn, same standard of proof.
Four boosted trees read the algorithm.
The system under observation is not a black box — it's a 42-logic machine across 20 engine groups. Between that machine and raw market history sits a fleet of gradient-boosted tree models including XGBoost: who earns where, whether this trade earns now, when pain cascades, and what a knob will forecast before silicon burns. Not “AI trading” — ML as infrastructure: the algorithm acts; the models decide what gets tested next.
Who makes money, where
XGBoost panel model: per logic × market regime → expected contribution. Which of the 42 logics earn in a 2004–2015 squeeze, which die, which only surface in modern tick regimes. The map retrains with the fleet and is served to the agents before they propose a change.
This trade, right now
Not “does the strategy earn” — “does this exact close earn.” López de Prado meta-labeling: P(win) and magnitude per extracted deal, trained on 20,522 real deals. Same pattern big-fund quants use — built on retail hardware. Proof-of-concept; labeled as such.
Drawdown early warning
Family concentration, unrealized pain, regime state → P(drawdown cascade over the next N bars). An early-warning classifier most retail stacks never run. In development — honest status beats a polished claim.
Forecast before the burn
Gradient-boosted trees: knob settings → profit / drawdown / profit-factor per range. Simulation cycles go only to ideas that forecast well. Dead-knob detection drops parameters the model cannot separate from baseline on any range — knobs that do nothing leave the search space.
One machine. A factory for agents.
One laptop is the whole physical stack. The rest is cloud and local LLMs anyone can rent — the difference is the factory floor: agents read the same bootstrap, follow the same laws, append to the same ledger, query the same knowledge base before acting. Cloud models reason; local models recall.
Not “agent generates code.” The pattern is research discipline at machine speed: form a hypothesis, burn cycles for evidence, accept or kill on the record, and pass institutional memory forward in a format designed for a future fine-tune.
A fund has departments — research, backtesting, validation. Here those are lanes with equal formalism: staged promotion, era-separated truth, multiple-testing care, go/no-go gates nothing — not even the operator — overrides alone.
Hypothesis → evidence
Agents propose one documented change, query lessons and the surrogate first, then earn the right to burn cycles.
Ledger over vibes
When intuition fights the append-only record, the record wins. Self-rejection is a feature, not a failure mode.
Successors inherit state
Knowledge base + ledger shape institutional memory — RAG for agents today, structured fuel for a fine-tune tomorrow.
Human + LLM = cyborg.
Not human slaving the LLM. Not human depending on the LLM.
The stance is efficient synergy: augmented over virtualized — human and model working together in pro of goals. The human keeps judgment, intent and the veto; the model multiplies reach, speed and pattern memory. Neither side is the slave.
Intent & veto
Goals, taste, ethics, the kill switch. The agent proposes; the human decides what ships.
Reach & memory
Pattern recall, parallel drudgery, tireless iteration — the boring 80% run without babysitting.
Cyborg loop
Human sets the bar, model hunts the delta, receipts come back for review. Compounding, not co-dependence.
RYIUK is the lab.
DAAVILE FX is where the pipeline gets battle-tested.
Same LLM + ML stack, first hostile domain: finance. What we flex for DAAVILE FX is not the trading — it's the sauce: an evidence ledger, multi-era truth tables, a staged robustness gauntlet, execution realism, an ML oracle, an agent fleet and a knowledge flywheel. Methodology is the moat. No signals, no courses — the process, documented.
Every claim has a run identity
Append-only record of real backtest runs. Deterministic run names; epoch rebases when the market-data base changes. Nothing is anecdote.
Regimes, not one lucky window
Hand-curated ranges across independent market eras — plus a completely untouched out-of-sample holdout no optimization has ever seen.
Staged promotion, fail = blocked
Clean → latency ladder → in-run Monte-Carlo stress → cost stress → forward. Champions are survivors, not declared winners.
Latency as a distribution
Strict profit, survival, ceiling and jitter gates per range. Ranked by probability-weighted expectation — never smeared into a single average.
The machine that reads the machine
Surrogates pre-rank candidates before a burn; feature importance over internal logics; dead-knob detection kills wasted sweeps; active learning picks the next teachable experiment.
A backtest factory
Parallel terminals through a typed MCP protocol, orchestrator manifests, campaign chunking, clean resume after mid-flight kills — data-center discipline, not hobby scripts.
Institutional memory that compounds
A structured knowledge base — internal logic map, per-range learnings, distilled lessons — RAG-queryable by the agents today, and explicitly shaped as training fuel for a future LLM fine-tune. Shared memory now; the model later.
Inquiries welcome.
Individuals and small sectors.
LLM workflows, pipelines, frameworks, fine-tuning, RAG, harness, inference, hardware, software — expand capabilities with top technology at reduced costs, protecting the wallet while the value stays solid.