RGI

Every AI system you buy today makes the same decision on Day 90 that it made on Day 1

Your people get smarter. Your tools don't. RGI changes that. Six decision loops compound on verified outcomes — scoring, evolution, situation analysis, factor trust all improve simultaneously. Your AI investment appreciates instead of depreciating.

Judgment stays when people leave System improves while you sleep Can't silently degrade — conservation law Knowledge transfers across domains Decision #1,000 is sharper than #1

pip install graph-attention-engine

v0.7.27 · 1,280 tests · Apache 2.0

LangChain · MCP · PyPI

↓ Demos & research below

RSI · Model Improvement

Retrain weights → ship new version

Context stays the same · Day 90 = Day 1

RGI · Context Improvement

Reshape graph from verified outcomes · model frozen

Context Graphs GAE / Math Decision Loops Judgment Memory

Same model · same code · smarter graph

Why now

The gap between the art of the possible and realizable results

$7–8T

projected AI infra spend
(Goldman Sachs, 2026)

90%

of firms report zero effect
on productivity (NBER, Feb 2026)

95%

of AI pilots report
zero return (MIT)

$2.18T

erased from KOSPI in 40 days
(Samsung/SK Hynix, Jul 2026)

Goldman Sachs, Jim Covello (May 2026): "Companies are losing more money implementing AI than two years ago. The hill that has to get climbed is even steeper today."

NBER (Feb 2026): ~90% of firms report no effect from AI on employment or productivity — yet executives project 1.4% productivity gains. The gap is between aspiration and measured outcome.

$184B

detected

$89B

analyzed · −$95B stalls

$23B

reaches execution · ~87% leaks

The insight-to-execution gap — where enterprise value actually leaks. Industry figures, modeled & illustrative.

$161B leaks at the decision layer

The one layer nobody optimizes. Until now.

For a $50M business: that ratio implies ~$4M in decisions that could compound but don't. For a solo trader: $16–26K/yr in leaks you can't see in a monthly P&L.

The "art of the possible" is settled. What isn't settled is the realizable art of the possible — and that's the whole game.

No system learns from what happened after

Models predict. The outcome is never written back. Decision #10,000 uses the same logic as #1.

Every vendor ships Day 1 capability. None compounds.

Better prompts, more parameters, bigger context windows. None of it makes decision #1,000 better than #1.

The gap is memory — verified, outcome-conditioned, governed

Not token memory. Not RAG. Decision memory that knows what worked, what didn't, and why — with a conservation law that prevents silent degradation.

The harness is the leverage, not the model

Harness design drives 6× more variation than model choice (Khattab 2026). RGI is six interlocking loops — the infrastructure that produces self-computing platforms, safe autonomous agents, and compounding intelligence.

RGI closes that gap — not by building a better model, but by making the decision infrastructure improve with every verified outcome.

The problem

Six hard problems that require six interlocking loops — not one better model

These aren't academic problems — they're where your budget goes to die. Decisions that don't improve. Patterns that don't transfer. Judgment that walks out the door. Every one is a line item your CFO can see.

①

No situation reasoning

Agents dispatch, they don't reason. 4% on copper in Q3 ≠ 4% on steel in Q1.

→ Situation Analyzer

②

Process changes are manual

Mining shows WHERE. Can't tell you WHY or fix it.

→ Tech-Process Fusion

③

Frozen production systems

Can't change the engine on a flying airplane.

→ AgentEvolver

④

No judgment geometry

Knowledge graphs store facts, not judgment.

→ GAE / Cross-Graph Attention

⑤

No safety envelope

No proof the system hasn't degraded.

→ Conservation Law

⑥

Every decision starts cold

RSI improves the model. RGI enriches the context.

→ Enrichment + DK weights

Full explainer with experiments →

How RGI solves it

Nine innovations. Six decision loops. The loops are the product.

What this produces: self-computing data platforms that learn which data matters. Autonomous agents bounded by a conservation law — safe to scale. Intelligence that compounds and transfers across domains. These capabilities didn't exist before because no infrastructure produced them.

①

Enrich

Context graph traversal

②

Decide / Abstain

Situation analysis → act or hold

③

Act

Execute the decision

④

Verify & Write

Outcomes reshape the graph

⑤

Evolve

AgentEvolver adapts at runtime

⑥ Reify ⑦ Discover ⑧ Govern ⑨ Compound

Compounds

Day 90 ≠ Day 1

Your decisions

Firm-specific, not generic

Discovers

Finds what nobody queried

Self-protects

Proves it's safe to scale

Non-transferable

Fork the code, start at zero

Open-source

Apache 2.0 · your machine

So what does this produce?

• A trader sees which setups earn trust and which cost money — $16–26K/yr in leaks surfaced
• A procurement team catches the supplier risk nobody queried — $680K/yr in prevented exceptions
• A data pipeline fixes itself and transfers the fix to the next pipeline — $1.62M/yr in auto-resolved bottlenecks
• A departed expert's judgment transfers to the next seat — ~844 values, not a handover doc

All modeled / illustrative. The demos show the reads. The numbers come from verified-outcome simulations.

The engine

Six mechanisms — each a recursive loop that compounds on verified outcomes

Why six and not one? Each mechanism solves one of the problems above. Together they produce what no single mechanism can: a system that gets measurably better while you sleep — and proves it's safe to let it.

TPF

Tech-Process Fusion

Celonis → WHERE · RGI → WHY + WHAT + LEARN + TRANSFER
→ $1.62M/yr: pipeline bottlenecks auto-resolved

Situation Analyzer

Reasons over context · decides or abstains
→ 4% on copper in Q3 ≠ 4% on steel in Q1 — context changes the decision

AgentEvolver

Evolves rules, configs & its own parameters at runtime
→ System improves between vendor releases, under safety gate

RL-CTRL

Matches enrichment effort to uncertainty
→ Hard queries get deeper investigation. Easy ones don't waste cycles.

Conservation Law

θ_min = 23.53/(α×V) · act vs abstain gate
→ System knows when it doesn't know — pauses before harm

GAE

Cross-Graph Attention

P(a) = softmax(−‖f − μ[c,a]‖²/τ) · ~194 experiments
→ Finds the risk nobody queried — cross-domain discovery

Landscape

What each platform does well — and where they stop

Palantir AIP

Strong ontology, agent tooling, government-grade deployment

Agents are deployed statically. Don't evolve from production outcomes. No situation analysis — execute workflows, not reason.

Celonis

Process mining leader — shows WHERE bottlenecks exist

Can't tell you WHY. Can't fix it. Can't learn from the fix. Can't transfer the pattern. RGI adds the other four verbs.

SAP Joule

1,300+ skills, deep ERP integration within the SAP ecosystem

Ecosystem-locked outside SAP. Agents don't learn from execution. Can't reason across SAP and non-SAP systems.

Snowflake / Databricks

Data gravity. Strong read-path infrastructure and catalogs

Read-path, not write-path. Can store context. Can't make it compound. Ontology-based enrichment — static map, same depth every query.

Zycus Merlin

5 procurement agents, 80%+ touchless AP, Gartner MQ Leader 2026

Generation 2: agents execute, don't compound. Invoice #10,000 = invoice #1. No conservation law. When the category manager leaves, Merlin is unchanged.

Agent memory (Mem0, Zep, Letta)

Solve statefulness for LLM agents — what happened, what's true, how to act

Memory without judgment geometry. Store facts, not verified outcome-conditioned centroids. No conservation gate. No compounding curve.

They optimize models. RGI optimizes the decision infrastructure. Every platform above can dispatch, monitor, or catalog. None writes verified outcomes back into a decision graph that compounds. Day 1 they look the same. Day 90 they don't. The question isn't which platform is best on Day 1. It's which one is better on Day 90 than it was on Day 1.

Why should you care

What six recursive loops produce after 90 days in a domain

This is what the $161B gap looks like in your business. Every number below is modeled from verified outcomes — not projected, not backtested.

Trading · Conservation

Scale capital without scaling risk

Before: act (unsafe) or keep humans in every loop (doesn't scale)

Now: conservation law proves $20K → $35K is safe — from YOUR execution data, not a backtest · 48% → 71% win rate

All Copilots · Judgment Memory

Judgment stays when people leave

Before: signal intuition, vendor knowledge, debugging patterns — walk out the door

Now: ~200–844 verified values per domain · Trading: $26K ahead in 6 months · Restaurant: patterns transfer in minutes

All Copilots · AgentEvolver

Improve between vendor releases

Before: frozen until v-next · thresholds, routing, SLAs all static

Now: detect drift → propose → shadow-test → promote/reject · three layers · under conservation gate

DataOps · Tech-Process Fusion

Fix it once, auto-resolve it forever

Before: Celonis shows the bottleneck · human investigates 4–6 hrs · next time: repeat

Now: WHY (schema change, 9× fanout) → WHAT (pre-join filter) → learn → transfer to billing_api · next time: auto-resolved · $1.62M/yr

S2P · Cross-Graph Attention

Find the risk nobody queried

Before: you find what you think to ask · supplier OTIF slips 96%→81% while price drops 8% — nobody connects the two

Now: cross-graph surfaces the correlation from 200 verified decisions · nobody configured it · the graph discovered it · $680K/yr

DataOps · Enrichment

Data that learns about itself

Before: ontology-based enrichment (Palantir, Databricks) — static map · same depth every query

Now: DK weights from verified decisions · source trust updates · combination discovery · investigation #1,000 traverses a richer graph than #1

Products

One engine, four domains

Trading

Your go-to setup is your worst — and you can't see it in a journal

48% → 71% win rate · $16–26K/yr in leaks surfaced (modeled, 400 decisions)

Demo Overview

S2P / Procurement

Who checks the auto-approved spend?

$680K/yr · 5:1 penalty (modeled, exception-rate simulation)

Demo Dashboard

Restaurant / Purchasing

Your best buyer's ordering patterns walk out when they quit

~844 judgment values · transfer to the next seat in minutes (modeled, 120-cell tensor)

Demo Dashboard

DataOps

The bottleneck isn't the join — it's a schema change

$1.62M/yr · 10:1 penalty (modeled, 9-pipeline simulation)

Demo Overview

The $184B question

Enterprise AI spending: $184B. Measurable return: $23B. A trader's journal shows P&L by month — RGI decomposes it per signal, per regime, per trade, and the scoring improves every month. A procurement system processes invoice #10,000 with the same logic as #1 — RGI's processes it with 10,000 verified outcomes sharpening every threshold. The industry optimizes models. RGI optimizes the decision infrastructure. That's not a better model. That's a new asset class.

Open-source · Apache 2.0Post-trade · observation-onlyYour data · your machine

Arindam Banerji, PhD

Founder · Dakshineshwari LLC

20+ years building enterprise AI platforms. Created and commercialized the Accenture Insights Platform (AIP) — one of the earliest full-stack AI/ML platforms. Creator of SOA; co-creator of e-Speak at HP Labs (now in the Smithsonian). Took the Analytics P&L at CSC/DxC from −26% margins to positive in 7 quarters. 9 US patents. PhD, Computer Science (Notre Dame).

Accenture · MD Cloud Analytics CSC/DxC · Global VP/GM Infosys · CTO Digital HP Labs 9 patents

RGI

The gap between the art of the possible and realizable results

Six hard problems that require six interlocking loops — not one better model

Nine innovations. Six decision loops. The loops are the product.

Six mechanisms — each a recursive loop that compounds on verified outcomes

What each platform does well — and where they stop

What six recursive loops produce after 90 days in a domain

One engine, four domains

Demos, overviews & research

Copilot Demos

Copilot Overviews

Engine

Research & Positioning