A²AIA2AI
What's Inside

Algorithms & models behind every tool.

Every tool combines a learned model, a curated SAP ontology and a reinforcement-learning feedback loop driven by SME verdicts. Scroll down for curated SAP, ERP and AI references.

Algorithm reference

The algorithms behind every A²AI tool.

Every tool combines a learned model, a curated SAP ontology and a reinforcement-learning feedback loop driven by SME verdicts. Below is the detailed pipeline for each of the seven core tools.

Tool · 01

Landscape Intelligence Pack

Model

Hybrid Named-Entity Recognition + regex extractors over BRD/TRD/RFP text; TF-IDF clustering for theme grouping; rule-based SAP module mapping.

Ontology

SAP business-process ontology (FI · CO · MM · SD · PP · QM · PM · EWM · TM · HCM · PS · BTP · MDG · IBP · EHS · CS) with synonym tables (e.g. ‘invoice’ → FI, ‘plant maintenance’ → PM). Custom-object classifier maps Z* / Y* patterns to RICEFW categories.

Pipeline
  1. 1. Parse uploaded artefacts (PDF · DOCX · TXT · CSV) → plain text
  2. 2. NER + regex extract modules, T-codes, CRs, custom fields, tables, roles
  3. 3. Cluster + de-duplicate; resolve synonyms via ontology
  4. 4. Emit Landscape Knowledge Base (KB) + downloadable Word / PDF status report
Reinforcement Learning

Each SME verdict (Verified / Accepted / Needs more info) is fed back as a reward signal to re-rank synonym confidence and downstream-tool boosts.

Tool · 02

Fiori App Recommender

Model

Lexical + token-overlap scoring with exact-match boosts for Fiori IDs and T-codes (regex /^(\/NS\/)?[A-Z][A-Z0-9_]{1,9}$/), plus ontology-based module routing.

Ontology

Role → Module ontology (16 modules). Each Fiori app is mapped to a functional module by inspecting role, name and product category against curated keyword regexes.

Pipeline
  1. 1. Read Landscape KB → detect modules in scope (boost +35)
  2. 2. Tokenise query, detect T-code / Fiori ID exact matches
  3. 3. Score every app: name 120 / role 40 / desc 25 / token overlap
  4. 4. Re-rank by ontology module boost, type and device filters
  5. 5. Present top-N with ontology trace (role + product → module)
Reinforcement Learning

Click-throughs and SME verdicts adjust per-module boost weights so the recommender learns which modules a given client truly cares about.

Tool · 03

Cost & Timeline Forecaster

Model

Quantile-regression LightGBM ensemble (P10 / P50 / P90) over historical SAP-programme telemetry, with Monte-Carlo overlay for risk distribution.

Ontology

Regional rate-card ontology (UK · EU · US · APAC), ACTIVATE phase ontology (Discover · Explore · Realize · Deploy · Run), RICEFW effort taxonomy.

Pipeline
  1. 1. Ingest Landscape KB + new inputs (region, modules, users, RICEFW)
  2. 2. Compute module-effort + scope add-ons × implementation multiplier
  3. 3. Add RICEFW days (R=4, I=12, C=15, E=8, F=6, W=10) and DM days (25/obj)
  4. 4. Derive team mix from composition preset, then duration from velocity
  5. 5. Apply rate-card, add licence + infra + training + contingency
  6. 6. Run Monte-Carlo → P10 / P90 envelope
Reinforcement Learning

Post-mortem actuals fed back to retrain quantile bands; SME ‘Needs more info’ flags raise uncertainty width on similar future projects.

Tool · 04

Clean Core Scorer

Model

Weighted XGBoost ensemble across 17 criteria in 4 dimensions (Extension 35% · Integration 25% · Process 25% · Data 15%) — aligned to SAP RISE Clean Core framework.

Ontology

SAP Extensibility ontology: On-Stack (Key-User / Developer Extensibility on ABAP Cloud) vs Side-by-Side (BTP). Integration ontology: Tier-1 Released APIs → Tier-2 → Classic RFC → Direct DB. Each criterion is a node in the ontology with a tier rank and a remediation path.

Pipeline
  1. 1. Read Landscape KB + new inputs; auto-fill where evidence exists
  2. 2. Score each criterion 0–10 against SAP Clean Core principles
  3. 3. Roll-up to dimension scores; weighted composite 0–100
  4. 4. Ontology resolves each ‘red’ score → standard remediation path
  5. 5. Output RISE-readiness tier + recommendations with ontology trace
Reinforcement Learning

SME accept/reject on each remediation tunes per-criterion weight and the strength of cross-criterion correlations.

Tool · 05

Change Impact Analyser

Model

Node2Vec graph embeddings over an 8,400-node SAP object graph + bidirectional BFS traversal for upstream/downstream impact; effort regressor per object type.

Ontology

SAP object ontology — Tables · Views · BAPIs · FMs · Transactions · Reports · Configs · BAdIs · Workflows. Each edge is typed (calls, reads, writes, configures, enhances).

Pipeline
  1. 1. Read Landscape KB; merge client-specific Z-objects into base graph
  2. 2. NLU → detect change anchor (table / FM / config) from free-text
  3. 3. Bidirectional BFS to depth 3 → L1, L2, L3 impacted objects
  4. 4. Effort regression per object × type (HIGH / MED / LOW)
  5. 5. Render force-directed graph + summary + ledger
Reinforcement Learning

Defect-leak signals from regression cycles feed back to re-weight edges and increase BFS depth for high-risk nodes.

Tool · 06

Test Coverage Intelligence

Model

Combinatorial test-design (pairwise + boundary value analysis) + scenario retrieval from a 3,200-case reference library; LLM-rerank for relevance.

Ontology

Test-type ontology (Happy · Negative · Edge · Integration · Performance · Security) × Module ontology (FI · CO · MM · SD · PP · QM · PM · WM · PS · BASIS · ABAP) × CR-pattern ontology (Master-Data · Transactional · Config · Interface).

Pipeline
  1. 1. Read Landscape KB + CR description
  2. 2. Match CR to ontology pattern → retrieve candidate library scenarios
  3. 3. Combinatorial generator fills gaps (pairwise / BVA)
  4. 4. Rank by risk-weighted coverage; group by test type
  5. 5. Emit executable test matrix + traceability to CR
Reinforcement Learning

Defects found in production map back to missing test types; the planner increases coverage of similar CR patterns in future runs.

Tool · 07

RICEFW Classifier

Model

XGBoost over TF-IDF features + 14 hand-crafted lexical features + naming-rule overrides for Z-namespace patterns.

Ontology

RICEFW ontology — Report (R) · Interface (I) · Conversion (C) · Enhancement (E) · Form (F) · Workflow (W). Each class carries an effort prior and a typical-tooling tag.

Pipeline
  1. 1. Read Landscape KB customs (or pasted inventory)
  2. 2. Tokenise + run rule pattern matchers (ALV, IDoc, LSMW, BAdI, …)
  3. 3. XGBoost ranks classes; naming-rule overrides finalise
  4. 4. Emit class + confidence + alternative classes
Reinforcement Learning

Mis-classifications corrected by SMEs feed straight into the rule library and become first-class features in the next training pass.

Curated references

The SAP, ERP and AI links our consultants actually use.