Six Months to Ship an AI Operations Secretary — A Solo-Dev Recap

The six-month map

To close the series’s first half: a month-by-month retrospective of Tasukiba Knowledge Relay’s six-month build.

2025-12: Kickoff (vision + tech stack)
   ↓
2026-01: Foundation (auth + multi-tenant)
   ↓
2026-02: Core (project + knowledge)
   ↓
2026-03: Suggestion engine (Voyage + pgvector)
   ↓
2026-04: Performance + billing
   ↓
2026-05: Finishing (quality gates + docs)
   ↓
2026-06-01: Release

Six months to MVP release. The distance from the problem in series part 1 to here, logged by month.

👉 Tasukiba product page

Month 1 (2025-12) — kickoff

Done:

• Articulating the vision (basis for docs/vision/README.md)
• Tech stack research
• Deciding on the name “Tasukiba”
• Prototype on Next.js + Prisma + Supabase

Learned:

• Writing the vision first — becomes the judgment criterion later
• Choose the tech stack early — changing it later hurts

Key ADRs: ADR-0001 (multi-tenant), ADR-0004 (Prisma)

Month 2 (2026-01) — foundation

Done:

• NextAuth v5 for auth
• Credentials + MFA (TOTP)
• Multi-tenant base design
• Service-layer authorization pattern

Learned:

• Auth: boring library is fine — NextAuth is enough
• Block tenant crossing from day one — bolting it on later is painful
• viewerTenantId as required argument is effective

Key ADRs: ADR-0005 (two-stage auth), ADR-0009 (MFA)

Month 3 (2026-02) — core

Done:

• Project CRUD
• Knowledge / risks / issues / retrospectives
• WBS / Gantt
• CSV import

Learned:

• Care about details in the core — UI consistency decides later experience
• Write tests as you go — writing them later is hell
• UX details like fullscreen dialog toggle matter

Key ADRs: ADR-0010 (state machine), ADR-0011 (soft delete), ADR-0017 (WBS import)

Month 4 (2026-03) — suggestion engine

Done:

• Voyage AI embedding integration
• pgvector + HNSW index
• 3-axis weighted score
• Graceful Degradation Mode

Learned:

• Semantic search is in a different class vs full-text
• Cost optimization — visibility='draft' doesn’t need embedding
• Fail-safe design — keeps running through external-API outages

Suggestion-engine internals: as written in B-2.

Key ADRs: ADR-0003 (suggestion engine), ADR-0008 (graceful degradation)

Month 5 (2026-04) — performance + billing

Done:

• Performance work (N+1 elimination, 1.8s → 0.7s, 60% improvement)
• Stripe Metered Billing integration
• ApiCallLog source-of-truth model
• DB / file capacity usage billing

Learned:

• N+1 elimination is powerful — simple fix, large effect
• Billing invariant is critical — root of business continuity
• 3-layer synchronization (current / cron snapshot / history queries) is necessary

Pricing detail: as written in B-3.

Key ADRs: ADR-0002 (per-call billing), ADR-0006 (Stripe), ADR-0020 (DB capacity)

Month 6 (2026-05) — finishing

Done:

• 5-stage quality gate complete (lint / tsc / test / coverage / build)
• E2E coverage enforced in CI
• Docs restructured
• Mascot introduced (as written in M-1)
• ADR-0019 billing redesign

Learned:

• Last month before release: finishing only — no new features
• Mascot introduction changes temperature — brand formation
• Doc restructuring boosts maintainability

Key ADRs: ADR-0014 (CRUD redesign), ADR-0019 (billing redesign), ADR-0021 (file capacity)

Cumulative stats

Total stats (2025-12 → 2026-05):
  - PRs: 450+
  - Lines of code: ~40,000
  - Tests: 1,500+ (unit) + 100+ (E2E)
  - ADRs: 21
  - KDD: 16,000 lines
  - Docs: 60+ files
  - Env vars: 30+
  - Tables: 30+

Solo, in six months. Honestly.

”Without AI-driven development, this wouldn’t have shipped”

Frankly: without AI assistance, six months to release would have been impossible.

• Code generation: ~70–80% by Claude Code
• Documentation: Claude drafts ADRs / KDD
• Review assistance: ask Claude to review

The feeling is that AI multiplies code-writing speed by 5–10x. Judgment with me, execution with AI — that division. A surprise: typing up “reproducible patterns” (in the sense of the psychological-safety post) gets dramatically easier when paired with AI.

Time investment

Alongside a day job:

• Weeknights: 1–2 hours × 5 = 10 hours/week
• Weekends: 4–6 hours × 2 = 10 hours/week
• Total: 20 hours/week

6 months ≈ 24 weeks = 480 hours.

That’s about 3 months of full-time development. Without AI assistance, easily 5–10x that.

What was hardest about solo development

1. Deciding everything alone

Design, business, legal — all alone. No reviewer.

Countermeasures:

• Make decisions visible via ADR
• Ask Claude Code “is this decision reasonable?”
• Let decisions sit for a few days before committing

2. Sustaining focus

Day job + indie development drains energy.

Countermeasures:

• Keep one day off per weekend
• When attention scatters, switch tasks
• Health first

3. The fear of releasing

“Will it actually work?” “Are there bugs?”

Countermeasures:

• Quality gates as structural guarantee
• Phased release (Dogfooding period)
• Emergency-response SOP ready

What I’d do again

• Write the vision first — judgment stays stable across 6 months
• Log ADR / KDD immediately — future-you is grateful
• Buy time with AI-driven development — solo 6-month MVP is achievable
• Set up quality gates early — once CI is in, later work is easier
• Dogfooding — using it yourself maintains quality (covered in L-2)

What I wouldn’t do again

• Over-featuring — should have stuck strictly to MVP-essential
• Premature optimization — some perf work could have waited until post-release
• Perfect UI — some details ate more time than they returned

The lesson: decide “what not to do” first matters as much as the to-do list.

The next six months

Post-release 6 months (2026-06 → 11):

• Accuracy tuning across features (chat semantic search / suggestion engine)
• Stripe live + initial tenants
• Dogfooding-driven improvements
• First teammate joining (planned)

Release was not the end — it’s a new start.

Connecting to year ten

The year-ten vision in L-3:

• “Cash cow” generating stable revenue
• Live proof of a psychologically safe organization
• Capital for multiple ventures

— the first step toward all of that is this six-month MVP. Small concrete results, connected to a long dream.

Summary

Month	Focus	Outcome
Dec	Kickoff	Vision + tech stack
Jan	Foundation	Auth + multi-tenant
Feb	Core	Project + knowledge
Mar	Suggestion engine	Voyage + pgvector
Apr	Performance + billing	60% improvement + Stripe
May	Finishing	Quality gates + mascot

Six months to MVP, recorded as a personal experience. Hopefully useful as realistic effort estimation for anyone considering indie SaaS.

First half of the series — wrapping up here (plus 2 technical bonus posts)

Thank you for staying with the first 20 posts of the Tasukiba Knowledge Relay series.

Chapter	Theme	Count
A	Origin & philosophy	6
B	Product overview & release	4
K	Community, hiring, culture	4
L	Post-release & roadmap	3
M	Mascot (bonus)	2
O	Six-month retrospective (this post)	1
Total		20

The series then closes with two technical bonus posts. Two of Tasukiba’s three unique features that haven’t been deep-dived yet:

• 6/15: Talking to Tasukiba in natural language — chat semantic search
• 6/16: Suggestions with a “Why” — LLM re-ranking

Coming over the next two days. The deeper technical writing goes to Qiita, but at the depth where homepage readers can grasp what kind of SaaS this is, these two posts are worth keeping here.

That’s the real first-half wrap. The second half continues based on how things actually unfold — Phase 2 / Phase 3 progress, Dogfooding insights, teammate-joining stories.

Tasukiba Knowledge Relay is always at the product page. Come say hi on Discord, anytime.

— Pass the tasuki forward.

Related posts

• Three master images for the mascot — series part 19, mascot ops
• An hour a day was evaporating — why I built Tasukiba — series part 1, the starting point of this six-month journey
• A “cash cow in ten years” — what I want Tasukiba to do in my life — where this six-month MVP is headed

About Tasukiba

Tasukiba Knowledge Relay is a business SaaS shipped on 2026-06-01 after six months of solo development. See the product page and give it a try.