PROJECT FACTS
CLIENT
Mid-market professional-services platform (multi-tenant)
SECTOR
Vertical AI tooling · multi-tenant SaaS
FACTA SERVICES
Fractional CAIO · SDD architecture · Spec authoring · Engineering delivery
TIMELINE
5 months (audit → spec → approval → production)
TEAM
1 fractional CAIO · 1 SDD architect · 3 engineers · 1 reviewer
STATUS
In production · 28 divisions + 4 cross-cutting modules live
Executive summary
One generic assistant becomes 28 governed, industry-specialized AI packages.
The client operated a multi-tenant platform where every customer got the same generic coding assistant. Sales tenants, finance tenants, healthcare tenants and legal tenants all shared one surface — no specialized agents, no division-scoped commands, no auditable doctrine. Adoption stalled exactly where the work got domain-specific: the assistant could write boilerplate but couldn't run a sales-pipeline review, draft a QBR, or close the monthly books.
FACTA deployed Package Builder — its spec-driven development (SDD) product — to industrialize the specialization. Every division was audited, gap-proposed, spec'd, approved, and only then implemented. In five months the client went from one generic copilot to 28 division packages and 4 cross-cutting modules, all namespaced under a single pb: plugin, portable across 14 supported coding tools, with CI guards green on every merge.
OUTCOMES
Divisions shipped
28
Plus 4 cross-cutting modules
Specialized agents
297
All lint-clean, division-tagged
Coding tools supported
14
claudecode canonical + 13 deltas
CI guards green
11
check-divisions, lint-agents, scan-secrets, etc.
The client
A multi-tenant platform where one assistant served every vertical.
The client runs a multi-tenant SaaS platform for professional-services firms — sales consultancies, finance teams, healthcare practices, legal firms, marketing agencies and operations groups all on the same product. Their AI surface was a single generic coding assistant shared across every tenant. The thesis was sound at seed stage: ship one copilot, learn what tenants actually reach for, then specialize. By Series A the learning had converged: 70% of usage was vertical work the generic assistant couldn't do well, and tenants were exporting prompts to run elsewhere. The platform needed to ship specialized packages per division without becoming 28 separate products.
The challenge
Specialize 28 ways without fragmenting into 28 products.
The mandate looked like a content problem and was actually a governance problem. Writing 28 sets of agents is easy; keeping them consistent, non-overlapping, tool-portable, secret-clean, and merge-safe across a growing engineering team is where the actual work lives. The client had tried a bottom-up approach — letting each division owner author their own agents and commands — and got overlap collisions, drift between tools, frontmatter inconsistencies, and a sprawling untested surface that no one could vouch for.
CONSTRAINTS
No overlap collisions
Two divisions authoring the same artifact (e.g. a BMC strategist, a retrospective facilitator) had to be detected and resolved before implementation — not in production.
Tool portability
Packages had to render on claudecode, cursor, codex, kilocode, gemini-cli, qwen-code, github-copilot, opencode, windsurf, aider, antigravity, kimi-code, openclaw, hermes, osaurus and mistral-vibe — without maintaining 14 separate codebases.
Approval gate before code
No division's implementation tasks could run until its spec was reviewed and approved. The client's regulators required an auditable spec-then-build trail.
Secret-free by construction
Every adapter and script had to pass scan-secrets.sh and route through ssrf_guard.py — no live-platform adapter could land without a 3-round security-reviewer gate.
FACTA's approach
Triple role: SDD architecture, spec authoring, engineering delivery.
FACTA operated as a single technical partner across the three planes the engagement required — direction, architecture, and execution — instead of three separate vendors:
01
Direction — Fractional CAIO
- ·Vertical roadmap across 28 divisions
- ·Priority order: red divisions activated before yellow depth-completion
- ·Tenant-scoped doctrine strategy (knowldage_database distillation)
- ·Investor narrative around specialized AI
02
Architecture — SDD design
- ·Audit → spec → approval → implementation gate
- ·pb: command namespace via Claude Code plugin manifest
- ·claudecode canonical + per-tool render-delta model (§9 of every spec)
- ·11 CI guards as no-regression backbone
03
Delivery — Engineering team leadership
- ·Led 3 engineers implementing spec §7 task-by-task
- ·Built OVERLAP-CHECK.md to resolve 8 cross-module collisions pre-implementation
- ·Distilled 43 knowldage_database works into traceable context packs
- ·Ran the approval review with the client's compliance reviewer
The technical solution
Package Builder: 28 divisions, 4 cross-cutting modules, one pb: namespace.
Package Builder organizes specialization into 32 modules — 28 industry divisions plus 4 cross-cutting modules — each with an audit, a claudecode/spec.md, and a per-tool render-delta block. Every spec follows the same 10 sections, every command is namespaced under /pb:, and every implementation task names the CI guards that must stay green. Implementation is forbidden until the spec is approved.
MODULES
DIVISIONS (28)
Industry packages — sales, finance, healthcare, legal, marketing, operations, governance, procurement, ESG, customer-success, and 18 more.
COMMAND-NAMESPACING
The pb: plugin manifest — every division consumes the colon prefix by reference rather than re-deriving it.
AGENT-ROSTER-GAPS
Cross-division roster audit — flags agents no division owns and proposes homes for them.
README-ATTRIBUTION
Provenance contract — adapted agents carry an Adapted-from line, KD distillations carry a Distilled-from line.
CHROME-TEMPLATE
Leak-guarded template surface — verify-template.sh + leak-guard.sh gate every templates/ edit.
OVERLAP-CHECK
The binding resolution table — 8 cross-module collisions resolved before any implementation task runs.
STACK
Open core · Spec-gated · 11 CI guards green on every merge
REPOSITORY
github.com/facta/package-builderResults
Five months. 32 modules. Zero overlap collisions in production.
Package Builder went to production with all 28 divisions and 4 cross-cutting modules live. The OVERLAP-CHECK table caught 8 cross-module collisions before any code ran — the BMC strategist, the retrospective facilitator, the ESG estimator, the incident-response command, and four others — each resolved with a binding owner and a re-scoped loser. The per-tool delta model kept 14 coding tools on one source of truth. And the approval gate produced an auditable spec-then-build trail that satisfied the client's compliance reviewer without slowing the engineering team.
METRICS
“Package Builder didn't just write us 28 packages — it gave us a way to keep them consistent. The overlap table alone saved us a quarter of arguing over who owns the BMC strategist. The approval gate is what let our compliance reviewer actually sign off.”
VP Engineering
Client platform
Replicable lessons
What makes specialization shippable instead of sprawling.
Specialization is a governance problem, not a content problem.
Writing 28 agent sets is easy. Keeping them non-overlapping, tool-portable, and merge-safe across a team is the actual work — and it needs an OVERLAP-CHECK and an approval gate, not more writers.
One canonical tool plus deltas beats N parallel codebases.
claudecode is the source of truth; every other tool is a §9 render delta. One spec, 14 tool surfaces — without the 14× maintenance cost.
An approval gate is a compliance asset, not a bottleneck.
Spec-then-build looked like overhead to the engineering team and like gold to the compliance reviewer. It's what let both sign off on the same shipment.
Namespacing by reference beats re-deriving per division.
Every division consumes the pb: prefix from command-namespacing instead of re-authoring it. One mechanism, 28 consumers, zero drift.
About FACTA
Fractional AI leadership with real technical execution.
FACTA is a consultancy specialized in fractional Chief AI Officer and CTO roles for companies that need to lead the transition to AI without hiring a full-time executive team. Unlike generalist consultants who deliver PowerPoint, FACTA delivers systems in production — including spec-driven products like Package Builder, where the spec is the approval unit and the implementation is the artifact.
We combine ten years of experience in AI architecture with operational expertise in multi-agent systems, spec-driven development, Graph RAG, A2A, and a modern stack (Rust, TypeScript, cloud and on-prem infrastructure). We work with post-seed startups through Series B scale-ups across Spanish- and English-speaking regions.
SERVICES
01 · Fractional CAIO
C-suite-level AI direction. Vertical roadmap, division priorities, investor relations, market positioning.
02 · SDD architecture
Audit → spec → approval → implementation gate. Module boundaries, overlap resolution, CI guard backbone.
03 · Spec authoring & engineering delivery
10-section specs with per-tool render deltas, implemented task-by-task with all CI guards green.
04 · Strategic Advisory
Supporting founders on fundraising, GTM, strategic partnerships and regulatory compliance.
A next case?
If you're building a multi-tenant platform that needs to specialize per vertical without fragmenting into N products, let's talk.
Book a 30-minute callhola@facta.dev