Agentic AI Research System

Running on Mac Mini, Docker and Tailscale

Architecture Evaluation

Cross-pipeline quality assessment · output scoring · comparative benchmarking · human review

Application tier

E2ER

End-to-End Research
24 workers · 118 skills · 16 stages

deployed

E2EREP

Empirical Pipeline
IV · DiD · RDD · panel data

deployed

E2ET

End-to-End Theory
60+ personas · 8 tiers · bisociation

pilot
Shared infrastructure

Knowledge Base

295K chunks · pgvector

Research Database

>90 GB · financial & blockchain · clear data governance

Evaluation Module

Quality gates · human review

Fetcher

Web access · injection guard

External data sources

MCP: Allium

DeFi · on-chain

Finance APIs

Traditional markets

arXiv / Sem. Scholar

Academic literature

FRED / WRDS

Macro & financial data


Architecture comparison

E2ER v1

Linear · 8 stages

Research inputquestion + scope
Literature fetchKB search + arXiv pull
Relevance filterscore & select sources
Summarize & extractclaims, methods, gaps
Analytical modelingquantitative analysis
Draft sectionprose generation
Review & revisecritique → rewrite
loop
OutputLaTeX / markdown
sequential · no parallelism

E2ER v2

Parallel workers + human gate

Research inputquestion + scope
↙ parallel fan-out ↘
concurrent workers
Fetcher Modeler Analyzer KB search
Merge & rankconsolidate outputs
Human review checkpointapprove / redirect / abort
Draft sectionprose generation
Review & revisecritique → rewrite
loop
OutputLaTeX / markdown
parallel workers · human gate · review loop

E2EREP

Causal-first · 6 stages

Paper inputtarget study + data
ID strategy auditIV / DiD / RDD validity
Data preparationclean, merge, panel build
Econometric analysisestimate + diagnostics
Robustness checksplacebo, subgroup, alt. spec.
Replication reporttables + verdict
structured · no open-ended search

E2ET

Divergent search · 6 stages

Research questiondomain + framing
↙ 60+ persona fan-out ↘
Cross-disc. searchper-persona lit. fetch
Bisociation enginecross-domain linking
Synthesis & rankingnovelty + feasibility score
Theory outputpropositions + agenda
max divergence → convergence