Five AI tools that automate the entire M&A lifecycle for Japanese mid-market deals: sourcing → due diligence → Day-1 readiness → 100-day PMI → cross-engagement knowledge layer.
The 4-constraint set (applied across the full portfolio — verified consistent across all 11 portfolio repos):
- Zero credit card — no paid API / cloud service required for the default path. A reviewer can clone, install, and run with $0 spend and no payment method on file.
- Local LLM (default) — when an LLM is involved, the default path is local (Ollama / similar) or deterministic mock. Paid cloud LLM is opt-in via env var, never default.
- Free / OSS only — every runtime dependency is permissively-licensed open source (MIT / Apache-2.0 / BSD-3); no proprietary SDK at build time.
- Security defense-in-depth — secrets-scan CI +
.gitignorehardening, encrypted-at-rest where PII is involved, append-only audit logging where applicable, dep-vuln gating (pip-audit/pnpm audit), paid-API constructor gate where applicable.
This meta-repo specifically demonstrates: the 5-tool MAIS suite hand-off chain (Citation array → IntegrationPlan → CockpitProject → PMICase) — each sibling carries the same 4-constraint baseline, with implementation details in their individual READMEs (see The five tools below).
Mid-market Japanese M&A advisory firms run the same five jobs on every deal: find candidates, run DD, plan Day-1, monitor 100-day PMI, transfer the lessons to the next deal. Each is partially automatable today; nobody has stitched the five into a coherent suite that respects Japanese mid-market specifics (family ownership, banking relationships, union dynamics) and Japan's 2026 amended Personal Information Protection Act.
MAIS is that suite, built end to end as a portfolio. Each tool is independently deployable; sibling tools hand off via versioned Pydantic schemas (Citation array → IntegrationPlan → CockpitProject → PMICase).
Two adjacent tool categories address parts of the M&A workflow in 2026, but neither delivers the five-stage automation chain MAIS demonstrates:
- Enterprise M&A platforms (DealRoom suite / Datasite Diligence / Devensoft / Midaxo) — integrated M&A workflow with deal pipeline + VDR + integration tracking, but priced for enterprise budgets and shaped around US/EU mid-market practice; Japanese mid-market specifics (family ownership, banking relationships, union dynamics, 2026 amended APPI) are not first-class.
- Generic AI document tools (ChatGPT for prompt-driven DD / Claude for clause extraction / Notion AI for knowledge management) — solve isolated pieces but require the human to chain them; no schema contract carries findings from sourcing through PMI handoff.
MAIS demonstrates: five tools sharing a versioned Pydantic schema chain, Japanese mid-market patterns first-class, 2026 amended APPI vault pattern, all running on a consumer laptop with zero credit card.
Target user: Japanese mid-market M&A advisory firms + PE-VC operating partners + corporate development teams + strategy consultancies running the full sourcing → DD → Day-1 → PMI → knowledge transfer cycle.
| # | Tool | Replaces / accelerates | Core tech |
|---|---|---|---|
| 1 | mais-deal-matching | Manual candidate-buyer matching by intermediaries | 5-stage hybrid retrieval (BM25 + dense + RRF + cross-encoder + LLM CoT) + PII Vault Pattern (SQLCipher, 2026 APPI compliant) |
| 2 | mais-dd-workbench | Manual VDR document hunting + DD question answering | Docling (IBM) ingestion + LlamaIndex CitationQueryEngine + CUAD/ACORD clause extraction + JP mid-market fit detector |
| 3 | mais-day1-cockpit | Manual Day-1 plan drafting + audience cascade | LangGraph 4-axis parallel agents + NetworkX dependency graph + 5-audience LLM communication kit |
| 4 | mais-pmi-cockpit | Manual KPI tracking + reactive next-action decisions | Apache Superset embed + Isolation Forest anomaly + LLM next-action recommender + multilingual sentiment |
| 5 | mais-pmi-knowledge-base | Tribal-lore knowledge transfer between deals | Self-built GraphRAG (NetworkX Louvain) + 5-axis weighted similarity + AI Assistant dialogue with citations |
Auto-rendered by docs/diagram/render.py (matplotlib + networkx, no network egress). The ASCII version below mirrors the same data flow for environments where images do not render:
mais-deal-matching mais-dd-workbench mais-day1-cockpit
(sourcing) ───► (DD) ───► (Day-1 plan)
│ │ │
▼ ▼ ▼
Profile/ Citation array, IntegrationPlan,
Company match JP fit pattern, PlanNode,
Q-A pairs RiskScore,
CommunicationKit
│
▼
mais-pmi-cockpit
(100-day)
│
▼
CockpitProject,
KpiSnapshot,
DriverInsight,
NextAction
│
▼
mais-pmi-knowledge-base
(knowledge layer)
│
▼
PMICase ADR,
Decision/Outcome,
Pattern,
AssistantQuery
Each tool publishes its outputs as Pydantic schemas; the next tool ingests via versioned mappers. Tools can be deployed independently or as a stack.
| Principle | How it shows up |
|---|---|
| Security-aware dependencies | Pinned versions in requirements*.txt; lock files committed where applicable; CI integration (pip-audit --strict, Dependabot, Dependency Review) is a deployment-time addition for client engagements rather than a portfolio-side claim |
| PII discipline | Vault Pattern (Fernet at rest + audit log append-only); embedding/matching never sees raw PII |
| No vendor lock-in | Anthropic API + OSS only; LLMProvider Protocol (3 methods) lets you swap to Ollama / Gemini / self-hosted in one file |
| Citation everywhere | Every LLM-generated answer carries source link-backs (page / cell / bbox metadata for Docling chunks) |
| Spec-Driven Workflow | Discovery → Requirements → Design → Tasks; module-boundary discipline enforced at code-review time; full drift-prevention infrastructure (GitHub Issues + Projects + linked PRs + CI checks) is added when an engagement begins |
| Consumer-laptop runnable | Every tool runs end-to-end on a single laptop with synthetic data; no GPU required for PoC scope |
| Synthetic data only | All tools ship with Faker (ja_JP)-driven synthetic data; no real PII / real engagements in any repo |
| Repo | Tests passing / collected |
|---|---|
| mais-deal-matching | 49 / 50 (1 integration test requires cache build) |
| mais-dd-workbench | 47 / 47 |
| mais-day1-cockpit | 66 / 66 |
| mais-pmi-cockpit | 96 / 96 |
| mais-pmi-knowledge-base | 245 / 248 (3 integration tests require eval data generation) |
| Total | 503 / 507 (4 integration tests require Quick Start data setup completion) |
Counts are AST-derived from tests/test_*.py files in each repo. To run:
cd <repo>
python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements-week4.txt # or week0/week1 depending on repo
pytest tests/Note: each repo has its own dependency closure; full install + run requires per-repo setup. Some heavy dependencies (multilingual-e5-large ~2.2 GB, cross-encoder ~134 MB) are downloaded on first index build.
Slate scale (~95% neutral) + warm amber accent (~5%) on off-white background. Golden-ratio (φ=1.618) typography scale. Designed for an executive aesthetic — clean enough to live behind a senior partner's screen during client meetings.
The two load-bearing cross-cutting decisions are recorded under docs/adr/ using the Nygard pattern (Context / Decision / Alternatives considered / Consequences / References):
- ADR-0001 — Suite shape: five tools, one schema chain, not a monolith — why five independently-deployable tools instead of a monolith / single deep tool / three-tool subset / 10+ tool sprawl
- ADR-0002 — Pydantic v2 schema chain as the inter-tool hand-off contract — why Pydantic over Avro / Protobuf / JSON Schema / TypedDict / in-process imports
Each tool's own docs/adr/ records the tool-internal decisions (stack choice, retrieval composition, vault pattern, etc.).
PoC portfolio. Each tool is functionally complete on synthetic data, with security-aware design choices documented per repo. For production deployment to a client engagement:
- Sandbox setup (Docker / WSL2 / Codespaces) for real PII / real documents
- Customer sandbox dry-run + 1-week stability before cutover
- External penetration test for large engagements
- LLMProvider swap to self-hosted (Ollama / vLLM) for data residency
- Per-paper license confirmation for real research ingestion (T5)
Commercial deployment is a separate engagement; contact via GitHub Issues.
- M&A advisory firms / intermediaries — internalize sourcing through PMI
- PE / VC funds — standardize portfolio company DD + integration
- Corporate M&A functions — deploy as internal AI tooling
- Strategy consultancies — offer to clients as a PMI advisory tool
MIT. See LICENSE. Each individual tool ships with the same MIT license.
This portfolio demonstrates capability. Production engagements operate under separate commercial agreements.