2025 R&D Journal¶
Overview¶
First year of R&D work on Rose - an agentic platform for inbound marketing.
Timeline: July - December 2025
Phases¶
Phase 1: Foundation (July-September)¶
Hypothesis: Monolithic prompt with standard RAG can handle B2B conversations.
- July: LangChain, LightRAG, TenantAwareNeo4JStorage
- August: Redis history, widget configuration
- September: MongoDB Atlas, reranking, metrics collection
Outcome: Infrastructure ceiling confirmed. Need specialized handling.
Evidence caveat for July–September. These three foundation months predate Rose's production telemetry and its dataset-evaluation harness (the harness was first built at the end of September; production analytics began in October). Their entries are grounded in the period's engineering record, but their result figures are qualitative pilot observations or pre-production estimates, not harness or production measurements. Read retrieval- and latency-quality numbers in Jul–Sep as direction, not measured metrics; the first measured quality and engagement numbers appear from October onward.
Phase 2: Visitor Intelligence (October-November)¶
Hypothesis: Understanding WHO visits improves conversion.
Rationale: September metrics showed technical questions convert 3x better than general questions. Without visitor identification, all conversations treated equally - missing opportunity to personalize for enterprise visitors.
- October: Visitor profiling, profile merge, dynamic questions
- November: 5-tier enrichment, interest signals, suggested answers
Outcome: +50% engagement, +58% conversation depth, +9% conversion.
Phase 3: Agentic Structure & Dialog State (December)¶
Hypothesis: Structured dialog state and parallel agent nodes improve conversation quality.
- December: Dialog state extraction, agent configuration, graph structure refinement
Outcome: Production-ready agentic architecture with 11 parallel nodes. Foundation for intent-based routing (Q1 2026).
Measured Business Impact (Q4 2025)¶
Initial Engagement Rate¶
Engaged Conversations (2+ turns)¶
Conversion Rate (Interaction to Form Submitted)¶
Summary¶
| Metric | Oct 2025 | Dec 2025 | Improvement |
|---|---|---|---|
| Initial Engagement | 1.24% | 1.86% | +50% |
| Engaged Conversations | 29.24% | 46.32% | +58% |
| Conversion Rate | 2.89% | 3.15% | +9% |
| Visitor Identification | 35% | 72% | +106% |
Monthly Entries¶
| Month | Main Focus | Key Hypothesis |
|---|---|---|
| July | Foundation & Architecture | Monolithic prompt + LightRAG |
| August | Production & Redis | Per-site configuration |
| September | MongoDB & Metrics | Systematic metrics collection |
| October | Visitor Profiling | LLM inference + enrichment |
| November | Enrichment & Signals | 5-tier pipeline + interest detection |
| December | Agentic Structure | Dialog state & parallel nodes |
R&D Innovations¶
| Component | Complexity | Status |
|---|---|---|
| TenantAwareNeo4JStorage | Very High | Production |
| Multi-Agent LangGraph | Very High | Production |
| 5-Tier Enrichment Pipeline | High | Production |
| Interest Signals Detection | High | Production |
| Intent Router | Medium | POC |
Bridge to 2026 R&D Projects¶
The 2025 journal predates the R&D project names introduced for 2026. Do not rewrite the 2025 monthly entries to use those names mechanically; instead, use the following bridge when reading 2025 work as the precursor to 2026 R&D projects:
| 2025 R&D line | 2026 R&D project | Continuity |
|---|---|---|
| Multi-Agent LangGraph, intent router, dialog state, action router | Adaptive Multi-Tenant Conversation Orchestration | The 2025 move from monolithic prompt to routed graph execution became the 2026 work on skill selection, deterministic execution contracts, and low-latency multi-tenant routing. |
| Tenant-aware storage, LightRAG, MongoDB checkpoints, visitor profiles, enrichment pipeline, interest signals | Compounding Context Engine for Company, Industry, and Buyer Intelligence | The 2025 work on retrieval, memory, profile merging, and enrichment became the compounding company, industry, and buyer context substrate. |
| Metrics collection, conversation labeling, business impact tracking, state serialization tests | Closed-Loop Agent Evaluation and Optimization | The 2025 measurement and regression foundations became the 2026 eval, observability, experiment-mart, and repair-loop work. |
This bridge is intentionally conservative: routine production hardening, client-specific onboarding, and product delivery remain non-R&D unless a monthly entry names a technical lock and experiment.
Where each 2026 project starts in 2025. The monthly entries now carry a one-line "Maps to 2026 R&D projects" pointer so the continuity is visible from inside each entry, not only here. The earliest month each 2026 project genuinely begins:
| 2026 R&D project | Starts | First work in 2025 |
|---|---|---|
| Compounding Context Engine for Company, Industry, and Buyer Intelligence | July | Multi-tenant graph-RAG isolation; hybrid graph+vector retrieval |
| Closed-Loop Agent Evaluation and Optimization | September | The dataset-evaluation harness (first measurement substrate) |
| Adaptive Multi-Tenant Conversation Orchestration | December | Intent classifier / deterministic action router decomposition (seeded by the September engagement experiment) |
| Synthesized Knowledge Publishing for Generative-Engine Discoverability | — | No 2025 work; this outbound-publishing line begins in 2026 |