🗺️ Phase roadmap: tech + business, with what's already done (companion to #83)

[teetangh]

Companion to #83 (business model decision of record). #83 says what we are building and for whom. This issue says in what order, and — just as important — what is already done.
Date: 2026-08-27 · Horizon: 12 months · Operating mode: solo founder, side project alongside a full-time job

---

How to read this

Two umbrellas, five phases each. Every phase has a Definition of Done — a single condition that must be true before the next phase starts.

• [x] = done and verified in the repo today
• [ ] = not done
• 🔴 = blocks the next phase · 🟠 = needed within the phase · ⚪ = nice to have
• 🅿️ = deliberately parked (see 🎯 Business model decision of record: B2B-first — sell seats and engagements to companies, not courses to students #83 Part 8)

Deliberately kept at capability level. Which library implements SSO or rate limiting is not a roadmap decision.

---

Where we actually stand today

Age	~8 weeks (first commit 2026-07-03)
Hand-written app code	33,403 LOC
Pages	77 across 8 route groups
Data models	111 (~34 have zero app code touching them)
Server-action modules	23
Unit tests	102 cases / 13 files ✅ passing in CI
End-to-end tests	14 cases / 2 specs ⚠️ written, not wired into CI
Migrations	4 (schema frozen behind a CI gate)
Open issues	55
Paying customers	0
Revenue collected	₹0
Live payments	No
Legal entity	Not registered

The honest read: the product is roughly 70% of the way to a first B2B sale. The company is roughly 10% of the way there. Almost every remaining blocker is on the business side, not the engineering side — which is the inverse of where the last 8 weeks went.

---

🛠️ UMBRELLA A — TECH

Phase T0 — Foundation ✅ COMPLETE

This is the part that's genuinely done. It is a lot, and it is good.

Identity & tenancy

• Multi-tenant model with distinct tenant types (creator / enterprise / university / hiring partner)
• Email + password and social sign-in, email verification, password reset
• Two-factor authentication
• Organizations, teams, dynamic roles, seat invitations
• Enterprise single sign-on paths (OIDC + SAML) with domain verification — built, never tested against a real identity provider (Enterprise SSO: verify OIDC E2E against real IdP + SAML/SCIM decision records #63)
• Tenant application → admin approval workflow
• Session storage with a secondary cache layer

Catalog & authoring

• Courses, sections, lessons, categories, resources
• Cohorts with capacity, enrolment windows and lifecycle states
• Creator studio: drag-and-drop curriculum builder
• Quiz authoring with automatic grading
• Assignment authoring with a grading queue
• Per-course and per-cohort pricing set by the seller
• Coupons, member management, earnings view
• Public creator storefronts

Learning experience

• Learner dashboard, lesson player, progress tracking
• Discussion threads per course
• Orders, billing and refund-request surfaces
• In-app notifications
• Community surface with leaderboard

The differentiated loop — mentor-reviewed projects

• Projects as a standalone sellable product across three tiers
• Milestones with review checkpoints
• Submissions with file upload to private storage (type + size validated)
• Mentor application → vetting → assignment workflow
• Versioned, immutable rubrics with per-criterion scoring
• Mentor review queue with SLA flags
• Milestone approve / request-changes and final review
• Completion gates actually enforced — every checkpoint approved, plus a passing defence review where required
• Credential issued on pass, with public verification page, printable certificate and public portfolio page

Money

• Cart and checkout with seat-capacity checks before payment
• Payment provider behind an abstraction, with signed webhooks
• Orders, payments, refunds, subscriptions, subscription plans
• Double-entry ledger — signed, append-only, idempotency-keyed, with cached balances
• Gapless sequential invoice numbering (GST requirement) via row-locked counter
• Commission snapshotted onto the order line at purchase time
• Mentor fee computed and posted to the ledger on pass
• Manual payout recording with reference capture

Enterprise / institutional

• Programs and program cohorts
• Organization licenses, seats, and credit pools with concurrency-safe drawdown
• Roster CSV import
• Report export
• Organization portal for the buyer

Platform & operations

• Admin console — tenants, mentors, projects, orders, refunds, payouts, coupons, licenses, moderation, feature flags, SSO
• Feature-flag infrastructure
• Audit logging
• Error monitoring
• Transactional email with a safe local fallback
• Terms, privacy and refund policy pages (text is solid; the entity behind them is a placeholder)
• Design system and marketing site
• CI: schema-freeze gate, schema-coverage audit, typecheck, lint, unit tests, build — with pinned actions and least-privilege permissions
• Deploy pipeline: PR previews, dev branch deploys, production branch

Definition of Done: ✅ met. A learner can be enrolled, learn, submit work, be reviewed against a rubric, and receive a verifiable credential — and the money movement is correctly recorded.

---

Phase T1 — Make it possible to take money 🔴 CURRENT PHASE

Nothing here is optional. Each item independently prevents a first rupee.

• 🔴 Go live on payments — currently test mode only · 🟠 Payments & money (Razorpay live) #38
• 🔴 Replace fabricated marketing content — the landing page's mentors, testimonials and statistics are hardcoded arrays. This is a compliance liability the moment a real buyer reads it · 🔴 Real content & brand (replace placeholders) #37
• 🔴 Real legal and support surface — support and grievance addresses are @example.com; legal entity reads "pending incorporation" · 🔴 Legal & compliance (India-first) #36
• 🔴 B2B tax invoices for contracts — a company will not pay without a compliant GST invoice · 🏢 B2B tax invoices for enterprise license contracts #46
• 🔴 Product analytics — without this the 90-day gate in 🎯 Business model decision of record: B2B-first — sell seats and engagements to companies, not courses to students #83 cannot be evaluated at all · PostHog product analytics + creator analytics dashboards #12
• 🟠 Fix free-course enrolment + coupon tenant-scope leak · 🔴 Free ₹0 courses unenrollable + coupon tenant-scope leak #58
• 🟠 Enforce entitlements beyond the player page; tighten the permissive catalog-scope default · 🔒 Enforce entitlements beyond the player page + tighten permissive catalogScope default #51
• 🟠 Enforce tenant-type gating at the route level (navigation-only today) · 🔒 Enforce TenantType gating at studio route level (nav-only today) #47
• 🟠 Security and auth hardening pass · 🔴 Security & auth hardening for production #35
• 🟠 Point the real domain at production — every link shared before this trains people on a URL that will change
• ⚪ Wire the existing 14 e2e tests into CI; add action rate-limiting · Launch hardening: action rate-limiting, prod email fail-closed, CI e2e job, DR docs, dead-code sweep #65
• ⚪ Remove shipped example/scaffolding routes

Definition of Done: a real company can be invoiced, can pay, and the payment lands in a real bank account — and we can see it happen in analytics.

---

Phase T2 — Deliver the first engagement 🟠

Only starts once T1 is done and a pilot is signed. Scope is driven by what the first buyer actually needs — do not pre-build.

• 🔴 Buyer-facing outcome report — completion plus rubric scores per learner, exportable. This is what the customer is actually paying for.
• 🟠 Verify enterprise SSO end-to-end against a real identity provider · Enterprise SSO: verify OIDC E2E against real IdP + SAML/SCIM decision records #63
• 🟠 Revoke credit-pool project instances on seat revoke / licence cancel · 🏢 Revoke credit-pool ProjectInstances on seat revoke / license cancel #48
• 🟠 Wire subscription-credit redemption — credits are granted today but never consumable · 💳 Wire SubscriptionCredit redemption — credits granted but never consumable #50
• 🟠 Handle subscription resumed → restore active status · 🔁 Handle subscription.resumed — restore ACTIVE status #49
• 🟠 Reconcile credit notes and ledger reversals when refunds fail · 💰 Reconcile credit notes + ledger reversals when Razorpay refunds fail #45
• 🟠 Content authoring gaps — sales description, thumbnails, richer editor, real free preview · Content authoring gaps: sales description, thumbnails/trailers, Tiptap editor, real free-preview, embed CSP #64
• ⚪ Clean up abandoned uploads · chore: cleanup abandoned UPLOADING MediaAssets (orphaned presigns) #68
• ⚪ Roster import dry-run and validation preview

Definition of Done: one cohort of real learners from one paying organisation completes, and the buyer receives a report they find credible.

---

Phase T3 — Make delivery repeatable ⚪

Only justified once there are 3 paying logos. Every item here is automating something currently done by hand — do it when the hand-work hurts, not before.

• Payout automation and creator/mentor KYC — payouts are manual today · Payout automation & creator KYC (RazorpayX) #14
• Notification infrastructure with WhatsApp-native reminders (review SLAs, deadlines) · Novu notification infrastructure + WhatsApp-native automation #10
• Bulk seat provisioning and self-serve org onboarding
• Auto-generated proof-of-work portfolio pages · Proof-of-Work portfolio pages (public, auto-generated) #6
• Search and recommendations upgrade · Search & recommendations upgrade + soft-bundle upsells #16
• Rubric calibration tooling — the quality risk that scales with volume
• Disaster-recovery documentation and restore drill · Launch hardening: action rate-limiting, prod email fail-closed, CI e2e job, DR docs, dead-code sweep #65

Definition of Done: a second and third customer can be delivered without the founder touching the pipeline manually.

---

Phase T4 — Scale surfaces ⚪

Nothing here before ₹1L/month recurring.

• Hiring-partner talent marketplace over verified portfolios · Hiring-partner talent marketplace (paid search over verified portfolios) #18
• SCORM/xAPI export for enterprise LMS interoperability · SCORM/xAPI export for enterprise interoperability #23
• Subdomain and custom-domain storefronts · Subdomain & custom-domain storefronts (Netlify automation) #13
• Mobile app evaluation · Mobile app — evaluate Fermion branded app vs Expo build #29
• Verification API for employers — the one genuine venture-scale door; pursue immediately if an employer ever asks unprompted

---

🅿️ Parked — explicitly not doing

#52 #53 #59 #60 #61 #62 #66 — video DRM, live sessions, live-join fallback, chat, sandbox/judge, vendor decision gate.

Rationale: async B2B delivery (content → assignment → rubric-scored human review → credential) needs none of it. This also retires the largest structural risk in the codebase — a stated differentiator depending entirely on a vendor that was acquired mid-transition, whose signup is paused, and for which we hold no API key.

Un-park trigger: a paying customer asks for live delivery in writing.

---

💼 UMBRELLA B — BUSINESS

Phase B0 — Research and positioning ✅ COMPLETE

• Full market research across Indian K-12/test-prep, global course marketplaces, cohort platforms, and solo-operator economics (~190 searches, 4 independent sweeps)
• Competitive teardown of direct comparables
• Audited-financial review of every named competitor — established that exactly one (Jaro Education) has positive PAT
• Unit-economics model with an explicit price floor
• Business model decision of record — B2B-first · 🎯 Business model decision of record: B2B-first — sell seats and engagements to companies, not courses to students #83 (pinned)
• Vendor risk register with a graded verdict
• Commission and payout economics designed and implemented
• Legal policy text drafted (terms, privacy, refunds)
• Brand name chosen, domain purchased
• Prior business viability review with kill criteria

Definition of Done: ✅ met. We know who the buyer is, what the price floor is, and which channels are arithmetically dead.

---

Phase B1 — Become a company that can invoice 🔴 CURRENT PHASE

This is the actual critical path. None of it is engineering. All of it is unglamorous.

Legal and financial

• 🔴 Register the legal entity
• 🔴 GST registration (18% applies to private training — it is not optional)
• 🔴 Business bank account
• 🔴 Payment gateway KYC and live keys
• 🟠 Real support and grievance contacts, published
• 🟠 Professional indemnity / basic business insurance check

Commercial paperwork

• 🔴 B2B pricing sheet — day rate, per-seat rate, engagement packages. Anchor at ₹50,000+ and validate against published market rates on the first three calls.
• 🔴 Standard services agreement / MSA template
• 🟠 Data processing addendum template (enterprise buyers will ask)
• 🟠 Mentor engagement agreement + payout terms
• ⚪ Purchase-order and procurement-form readiness

Positioning assets

• 🔴 One-page offer — what we sell, to whom, at what price, with what outcome. One page, not a deck.
• 🟠 Short capability deck for L&D buyers
• 🟠 Case-study template, ready to fill the moment a pilot completes

Definition of Done: we can legally issue a compliant invoice to an Indian company and receive money for it.

---

Phase B2 — First paying customer 🔴 (the 90-day gate — target 2026-11-25)

Choose the beachhead — one, not four

• 🔴 Pick a single vertical and commit for 90 days. Ranked by evidence of budget, not interest:
  1. GCC engineering upskilling — 2,100+ centres, 2.36M professionals, stated Dodo Payments provider — Merchant of Record for international sales #1 problem is "finding the right talent," explicit preference for precise skill fit over volume. Strongest evidence.
  2. AI/ML for working engineers — 76.6% of Indian orgs prioritising it; crowded, so verticalise by role or industry
  3. Semiconductor/VLSI specialisations — government-identified gap in non-generic specialisations; too small for the giants ⚠️ price by phone; no public fee data exists
  4. Tier-2/3 college placement prep — real gap, but slow-paying and seasonal

Build the supply bench (target 5, cap 10)

• 🔴 Recruit 5 practising senior engineers who can review async against a rubric
• 🟠 Calibrate them on a shared rubric — run 2 sample reviews each and compare scores
• 🟠 Agree payout terms and SLA expectations in writing

Sell

• 🔴 Build a named target list of 50 companies/institutions
• 🔴 25 outreach conversations via LinkedIn + WhatsApp + warm intros
• 🔴 5 qualified discovery calls
• 🔴 1 pilot signed — free or discounted, explicitly in exchange for a public case study and a named reference
• 🔴 1 paid engagement ≥ ₹50,000, invoiced and collected
• 🟠 Written loss reasons captured from everyone who said no

Discipline

• 🔴 ₹0 on Google/Meta/Instagram ads. At ₹1,500–4,000 CAC against products we keep 20% of, small budgets are a pure leak. Revisit only at AOV ≥ ₹40,000 or for warm retargeting.
• 🔴 Gate review on 2026-11-25 against 🎯 Business model decision of record: B2B-first — sell seats and engagements to companies, not courses to students #83 Part 7

Definition of Done — PASS if ANY: ≥1 paid corporate/GCC engagement collected · ≥1 institution on a paid seat licence · ≥3 paying B2B logos with <8% refunds.
FAIL all three → stop feature work entirely, two-week diagnosis interviewing everyone who said no, then narrow to whichever segment showed warmth or make the deliberate call to treat this as a portfolio asset. An honest stop beats a slow fade.

---

Phase B3 — Make it repeatable 🟠 (months 4–6)

• Publish the first case study with real numbers — no invented statistics, ever again
• Reach 3 paying logos
• Raise the price on customer Live defense engine — AI-assisted project defense (AVNI-style) #3 and see if it still closes
• Secure 2 referenceable customers willing to take a call
• Document the sales motion: what the pitch is, what objections come up, what closes
• Grow the mentor bench to ~10 with calibrated rubric scoring
• Measure, don't estimate: contribution per engagement, effective CAC, refund rate, review SLA breach rate
• Land one second engagement from an existing logo without discounting — the real signal
• Decide whether colleges are worth the slow-payment cycle, on evidence

Definition of Done: a repeatable, written sales motion that has closed three times, with measured — not projected — economics.

---

Phase B4 — Scale ⚪ (months 7–12)

• Reach ₹1L/month — roughly 2 billable days, or 90 enterprise seats, or one high-ticket cohort seat every six weeks
• Open a second vertical only after the first is repeatable
• Explicit fork in the road: stay solo · make a first hire · raise capital
  • ⚠️ Indian edtech funding was $118M–$249M across 22–31 deals in 2025, an 8-year low. Assume no institutional capital. Plan to be default-alive.
• Pursue a verification/API conversation only if an employer initiates it
• Consider productising the assessment layer once ~10 logos exist

Definition of Done: revenue covers the founder's time at market rate, or a deliberate decision has been made not to pursue that.

---

🔗 Where the two umbrellas block each other

Business step	Blocked by tech	Tech step	Blocked by business
Issue any invoice	#36 entity surface, #46 B2B invoices	Go live on payments (#38)	Entity + bank account + gateway KYC
Send any prospect a link	#37 real content, domain cutover	Buyer outcome report (T2)	Knowing what the first buyer wants
Evaluate the 90-day gate	#12 analytics	Everything in T2	A signed pilot
Close an enterprise deal	#63 SSO, #35 security	Payout automation (#14)	Enough mentors to make it hurt

The single most important line in this table: going live on payments is blocked by company registration, not by code. Registering the entity is currently the highest-leverage hour available, and it cannot be done in an editor.

---

📅 Indicative sequence

Window	Tech	Business
Weeks 1–4	T1 blockers: analytics, real content, entity surface, B2B invoices	B1: register entity, GST, bank, gateway KYC, pricing sheet, one-page offer
Weeks 5–8	Finish T1: live payments, entitlements, security, domain	B2: recruit 5 mentors, build target list of 50, start outreach
Weeks 9–13	T2 only as the pilot demands	B2: discovery calls → pilot signed → first paid engagement
Gate: 2026-11-25	—	Evaluate #83 Part 7. Pass, narrow, or stop.
Months 4–6	T2 completion, selective T3	B3: case study, 3 logos, raise price, measure
Months 7–12	T3 automation	B4: ₹1L/month, second vertical, the fork

---

🚨 Standing tripwires (from #83 — repeated because they matter)

• Any rupee on paid ads before AOV ≥ ₹40,000 → stop, revert to outreach
• Mentor review SLA breach >20% for two consecutive weeks → pause all selling, fix supply first
• Refund rate >12% → freeze acquisition, re-price
• New feature work >25% of founder hours while zero gate items are met → scope quarantine
• Two consecutive months with <5 qualified B2B conversations → the motion isn't real; re-target

---

The uncomfortable summary

Eight weeks produced 111 data models, 77 pages, 33,000 lines of well-tested code, and zero customer conversations. The engineering is not the constraint and never was.

Phase T0 is genuinely impressive work. Phase B1 is four government forms and a bank appointment. B1 is worth more than T2, T3 and T4 combined right now, and it is the phase most likely to be postponed — because it produces no diff.

The next artifact should be a signed pilot, not a merged PR.

[teetangh]

📚 Reconciliation: the two spec documents in docs/ vs. this roadmap

Both Word documents in docs/ have now been read in full and reconciled against #83 and this issue. They were not part of the original codebase audit — their existence was noted, their contents were never opened. This comment closes that gap.

Document	Size	Dated
docs/Product_Requirements_System_Design_Spec.docx	~50,600 chars · 14 sections	July 2026, "Founder Working Draft v0.1"
docs/Competitive_Teardown_Euron_Systems_Crio_do.docx	~33,900 chars · 8 sections	July 2026, "Working Draft Vol. 2"

Verdict: ~40% still load-bearing · ~35% factually stale · ~25% actively harmful if followed today.

These are genuinely good documents — sourced, opinionated, with explicit recommendations on every open question and an honest "where to concede" section in the battlecards. Nothing below is a criticism of their quality. They are strong answers to "what should we build?". They simply never ask "who will pay, how much, and have we asked them?" — which is the question #83 was forced to answer.

⚠️ Neither file carries any indication that it is partly superseded. Anyone opening docs/ cold — including future-you — will read §13's roadmap as current. It is not.

---

✅ STILL BINDING — do not re-litigate these

PRD §4 — Build vs. Buy. Vindicated by events.
The recommendation was: rent the commodity layer, own the differentiated layer, and keep vendor integrations behind your own service boundary so the engine is swappable without a rebuild. It named the exact failure mode — "the real danger of an all-in-one vendor is designing your own data model and workflows around its assumptions."

That advice was followed, and it is the sole reason the video/live/labs vendor being acquired mid-transition is a parked issue rather than an existential one. The document predicted the risk that materialised and the mitigation held. This is the single best call in either doc.

PRD §9 — System design, module responsibilities, conceptual data model.
Technology-agnostic and an accurate description of what was actually built. Still usable as architecture documentation today. No changes needed.

PRD §9.6 — Non-functional requirements. Now more relevant, not less.
Two items in particular are worth more under a B2B strategy than they were under a consumer one:

• "Integrity of mentor-review data — reviews, rubrics and defense records are effectively the platform's credibility asset and should be treated with the durability and auditability of a compliance record, not disposable application data."
• DPDP / GDPR-class handling for institutional learner data, called out as "a real gating factor in winning institutional contracts." Correct. This is Phase B1 paperwork (DPA template), not a nice-to-have.

Teardown §5.2 — the entire SKIP list. Every item still holds.

• ❌ Big up-front placement guarantee paired with rigid, contested refund terms — identified as "the specific, recurring mechanism behind Crio's public trust damage." Directly reinforces the 🎯 Business model decision of record: B2B-first — sell seats and engagements to companies, not courses to students #83 call to kill Income-share / success-fee option — fair terms designed FIRST #25 (income-share/success-fee).
• ❌ EMI partnerships without hardship or pause provisions — consumer-complaint evidence of real harm.
• ❌ Breadth-over-depth catalog without a quality bar per course.
• ❌ Inflated feature-count marketing ("200+ AI agents").
• ❌ Commission-driven outbound sales-agent motion.

Teardown §7 — one sequencing call that still stands: build the fair refund/guarantee design before building any guarantee at all. Cheap to build now, expensive to retrofit after a trust incident.

---

⬆️ PROMOTED — buried in the docs, now the core motion

PRD §6.4 — Enterprise & University Programs.

This is the most valuable section in either document and it was filed as pillar 4 of 4, scheduled at V2 (12–24 months) in §13.

It contains real primary research into how IIT and IIM programs actually operate — reviewing published partnership structures from IIT Jammu, IIT Madras (CODE/NPTEL) and IIM Tiruchirappalli — and concludes that the consistent pattern is: the institution keeps academic authority and certification; an empanelled partner handles reach, learner support and operations. From which it draws the right inference:

"That is an operating relationship, not a software license — the enterprise sales motion needs to sell delivery capability, not just seats."

That is the #83 thesis, reached independently a month earlier, and then deprioritised to position 4 of 4. It also correctly identifies the gap: available partners today are "either broad generalist upskilling companies with limited hands-on technical depth, or narrow LMS vendors with no operations muscle."

Action: §6.4 is now Phase B2 material, not V2. Reread it before the first institutional pitch — the empanelled-partner framing is directly usable in a deck.

PRD §6.4 also pre-answers the packaging question: don't force a choice between courses and projects — sell a configurable Program (curated course sequence + optional mentor-guided capstone + co-branded certificate + cohort ops/reporting). That is exactly the SKU to price in Phase B1.

---

🕰️ FACTUALLY STALE — several anchors were already wrong when written

The docs are dated July 2026 and carry an honest caveat to re-verify price-sensitive claims. Taking them up on it:

Claim in the docs	Status as of 2026-08-27
Udemy treated throughout as a live independent competitor with a "Personal Plan"	Coursera's acquisition closed 11 May 2026 — two months before the doc was written. Udemy is delisted.
Crio's Skill-Lync acquisition framed as "real institutional backing, not founder-only resources"	Skill-Lync did 400+ layoffs (Apr 2023) and 225 more (Jul 2023), closing Delhi NCR, Mumbai and Pune offices. That was distress, not backing.
Fermion recommended as the pilot infrastructure vendor (§4.3)	Acquired by Testpress; signup paused; no API key obtainable. Parked — #66.
Scaler cited as a high-ticket benchmark to price against	FY25 revenue ₹363 Cr, −5.5% — shrinking, near-breakeven only after cutting salaries 26.5%.
upGrad cited at ₹99K–₹3L+ programs as the incumbent to beat	FY26 revenue ₹2,070 Cr (+7%), net loss ₹130 Cr.
Maven's "~$500 avg course, 10% platform fee" cited approvingly as pricing precedent	Maven's own published figure: average Year-1 instructor earnings $10,000, and ~80% of demand is instructor-supplied, not marketplace-supplied.
Euron Systems positioned as a live infra competitor to bake off against	Still worth a look, but the bake-off (#66) is parked along with the rest of the vendor track.

Nothing here is a research failure — competitor facts in this sector decay in weeks. It is a reason not to price or pitch off these tables without re-verifying.

---

🛑 SUPERSEDED — do not follow these sections

PRD §3.3 — the six-sided marketplace framing.

"The clearest framing: this is a six-sided marketplace, not a single product."

Presented as the core strategic insight. It is the direct ancestor of 111 data models (34 with zero app code), 77 pages, and zero customers. It authorised building every side simultaneously, which is precisely what happened.

The doc's own §12 lists "two-sided marketplace cold start" as risk 6 of 6. It was risk #1, and it is the one that materialised. #83 replaces this framing entirely: one buyer type, one motion, supply hand-recruited to ~5.

PRD §8.2 — the pricing ladder (Free → Learn → Mentor → Career → Enterprise).
A consumer subscription architecture. Superseded by the ₹50,000 price floor and a corporate buyer. The tier-naming analysis is thoughtful and now moot — we are not selling consumer subscriptions.

PRD §8.3 — à la carte project pricing (Sprint ₹2,500–6,000 · Capstone ₹12,000–30,000 · Flagship ₹35,000–90,000).
Only the Flagship band clears the floor. Sprint at ₹2,500 is inside the zone the research identifies as unwinnable without a years-old free-content funnel.

PRD §13 — the phased roadmap. ⚠️ The most dangerous section to follow.
It schedules the creator marketplace and self-serve creator onboarding at MVP/V1, and enterprise Program packaging at V2 (12–24 months). This issue inverts that. If you open docs/ looking for sequencing, use #84 instead.

PRD §11 — success metrics. GMV, creator retention, storefront conversion, mentor utilisation. Marketplace metrics for a business with no marketplace. Note what is absent: no revenue-per-customer target, no CAC target, no LTV:CAC ratio, and no "have we asked anyone to pay yet" milestone anywhere in the document.

Teardown §5.3 — the "BUILD EXTRA / genuine whitespace" list.
The whitespace claims are accurate — neither competitor covers à la carte projects, multi-creator discovery, or project-authorship verification as a headline trust claim. But whitespace is not demand. Part of why nobody occupies that space is that nobody pays for it. Specifically, "project-authorship / anti-AI-shortcut verification as a headline trust claim" runs directly into the HBS + Burning Glass finding in #83 §9.1 — employers changed hiring for ~1 in 700 hires (0.14%). Keep the capability; do not lead with it.

Teardown §5.1 — the STEAL list is mostly sound but now mistimed. The Crio-style free staged funnel (10,000+ applicants/cohort, quiz-gated stages) is the best acquisition idea in either document — and it is a B2C engine. Correctly deferred as #7 under a B2B strategy. Revisit only if a consumer motion is ever reopened.

---

📌 Net effect on this roadmap

No changes. Phases T0–T4 and B0–B4 stand as written. The docs supply supporting material, not a competing plan:

• Phase B1 — pull the Program packaging definition from PRD §6.4; pull the DPA/data-protection requirement from §9.6; pull the fair-refund-window design from Teardown §7.
• Phase B2 — the IIT/IIM empanelled-partner research in §6.4 is directly usable in the first institutional pitch.
• Phase T2 — PRD §9.2 module responsibilities remain accurate documentation for the buyer-facing outcome report.
• Ignore — §3.3, §8.2, §8.3, §11, §13, and every competitor price table without re-verification.

🔧 Suggested housekeeping (not yet done)

Add a short "partially superseded — see #83 and #84" header to the top of both documents, or an equivalent note in docs/README. One commit. Prevents §13 being followed by accident six months from now, when the reasoning in this thread is no longer in anyone's head.

---

One observation worth recording. The PRD was written before the market research and asks what to build. The research asked who pays. Both documents are strong on the first question and silent on the second — and the eight weeks that followed produced 33,000 lines of well-tested code and zero customer conversations. That is not a coincidence; it is what happens when the planning artifact has no revenue milestone in it.

[teetangh]

🧮 Stress-testing the solo-operator thesis: "low cost base + vibe-coded platform = recurring money plant"

Recording this because it is the load-bearing assumption behind everything else, and it is half right in a way that matters and half wrong in a way that is fatal. Worth writing down before it gets tested with real months.

The thesis, stated fairly:

The giants lose crores because their cost base is enormous. Mine is one Claude subscription and some monthly SaaS. I can build the whole platform quickly, it then runs indefinitely, teachers arrive and teach, senior engineers mentor projects, and I take a commission tuned to be attractive to instructors and still profitable for me. One-time build cost, recurring revenue.

---

✅ WHAT IS ACTUALLY RIGHT — and it is not a small thing

1. The cost-base advantage is real, and it is the single strongest card in hand.
PhysicsWallah needs ₹3,900 Cr of revenue to roughly break even. upGrad needs ₹2,070 Cr and still loses ₹130 Cr. This operation needs ₹30–60 lakh. Those are not the same game, and analysis calibrated to the first is genuinely the wrong lens for the second. Point conceded.

2. Not needing venture returns is a structural advantage, not a consolation.
Scaler reached near-breakeven in FY25 only by cutting salaries 26.5% and marketing 30% — and revenue shrank 5.5% as a result. They are trapped: the growth that justifies their valuation requires spend that destroys their margin. That trap does not exist here. A business that is allowed to be small can make decisions a funded business cannot.

3. Small, profitable operators demonstrably coexist with the giants.
Not theory — filings: Learnyst (bootstrapped, ~₹8.17 Cr, 12 years, 42 people) · Academically Global (₹30 Cr, bootstrapped, healthcare licensure) · Jaro Education (₹52.9 Cr PAT on ₹273.9 Cr) · iMocha (profitable before raising) · Taro ($216K ARR, solo, bootstrapped). The category is not closed to a solo operator.

4. Variable-cost supply is the correct instinct. Paying mentors a share per review instead of salaries means supply cost scales with revenue instead of preceding it. That is exactly right and it is already how the ledger is built.

So the conclusion is not "a solo operator cannot win here." It is: the win does not come from where this thesis assumes it does.

---

❌ WHAT IS WRONG — three things, in order of severity

1. The cost that killed those companies was never engineering

This is the crux. The thesis assumes their losses came from building and running the platform. Their own filings say otherwise:

Company	Fiscal year	People	Sales & marketing	Combined as % of total expenses
upGrad	FY25	₹703.6 Cr	₹303.6 Cr	~52% of ₹1,942.6 Cr
Simplilearn	FY25	₹187 Cr	₹134 Cr	~52% of ₹621 Cr
Vedantu	FY25	₹219 Cr	₹27 Cr	~55% of ₹444 Cr
Eruditus	FY24	—	₹1,007 Cr	marketing alone

Not one of those lines is servers, or engineers building an LMS. They are salespeople, teachers, counsellors, and advertising — the cost of acquiring customers and the cost of delivering to them with humans.

Eliminating engineering cost removes a line item that was never the binding constraint. The thesis inherits 100% of the customer-acquisition problem and 0% of the customer-acquisition capability. upGrad burns ₹303 Cr a year on marketing and still only grows 7% — that is what it costs them to find buyers, with a brand and a sales floor. The plan here is to find those buyers with ₹0 and evenings.

That can work — but only via a channel that costs nothing and does not scale linearly, which is exactly why #83 lands on institutional buyers (one conversation delivers 20–300 learners) rather than a self-serve marketplace.

2. "Teachers will come" is the specific assumption that has never held

Every platform that made this bet, at every funding level:

Platform	Resources	Outcome
Maven	$30M, a16z, 5 yrs, ex-Udemy co-founder, 10% take	~20% of instructor sales come from Maven. The other 80% instructors bring themselves. Published avg Year-1 instructor earnings: $10,000.
Classplus	~$159M	Tutors reverted to WhatsApp and YouTube. Pivoted to being a test-prep company.
Topmate	300,000+ creators onboarded	~$1M cumulative creator earnings — roughly $3 per creator, lifetime.
FrontRow	$17M	Died at $3–4M ARR, marketing exceeding 100% of revenue.

The asymmetry that breaks the model:

• A teacher with an audience does not need a platform. They already have Razorpay + YouTube + WhatsApp, or a 5%-fee tool. Giving them 80% of a sale they generated themselves is worse than giving them 100% on their own site. What exactly is the 20% buying them?
• A teacher without an audience is worth nothing to the platform, because they bring no students.

A marketplace's product is distribution. This platform has none — 0 users, 0 traffic, an unindexed domain. Onboarding instructors is selling a service that does not yet exist. That is the cold-start problem, and it is not solved by better software; Maven proved that with $30M.

3. "One-time cost, then it runs forever" is wrong on two axes

(a) Software is a depreciating liability, not a fixed asset.
111 data models, 77 pages, 4 migrations, a payment integration, a tax-compliant invoice sequence. That surface requires: dependency and security patching, payment-gateway API changes, GST/TDS rule changes, framework major versions, expiring certificates and webhook secrets. Unattended software does not "run forever" — it rots quietly and then fails at the worst possible moment, which in a payments system is while holding someone else's money. ~34 of the 111 models have no app code at all — that is pure maintenance surface generating zero revenue.

(b) The real recurring cost is human, and it is yours.
A marketplace is an operations business wearing a software costume. The recurring work is:

• Payouts are manual today (Payout automation & creator KYC (RazorpayX) #14) — an admin types an amount and a reference by hand, per instructor, per mentor, per cycle
• Chasing mentors on review SLAs — the Dodo Payments provider — Merchant of Record for international sales #1 quality risk flagged in the roadmap
• Refunds, chargebacks, disputes, ledger reconciliation when a payment half-fails
• Vetting instructors and content before they damage the brand
• Support: "I paid and can't access it" at 11pm
• GST filing, and TDS obligations — facilitating third-party sales likely makes this an e-commerce operator under §194-O, with 1% TDS deduction on seller gross sales. That is a recurring statutory duty, not a one-off.
• One bad instructor, one plagiarised course, one student who paid and got nothing → you are legally and reputationally liable, not them.

None of this is automatable at low volume, and all of it is interrupt-driven — which collides directly with a full-time job. "Let it run forever" describes a product. This is an operation.

---

🔢 The commission model, worked

At a 20% take, to earn ₹1L/month the platform must clear ₹5L/month GMV — about 100 sales/month at ₹5,000, every month, from zero traffic.

And the instructor's side of that: they earn ₹4L for supplying 100 students they had to find themselves. At which point the rational move is to keep 100% on their own storefront.

Against the B2B alternative in #83: two billable days a month = ₹1L.

100 consumer transactions a month, forever, versus 2 conversations a month. Same revenue. One of them is compatible with having a job.

---

⚠️ The uncomfortable one: cheap building is not an advantage anymore

The thesis rests on being able to build fast and cheaply with AI. That is true. It is also true for everyone else.

When building a marketplace drops from $2M and 18 months to one person and 8 weeks, software stops being the moat — it becomes the commodity. What stays scarce is precisely what cannot be generated: a signed contract, a mentor who actually reviews on time, an institution that trusts an unknown vendor with its students, a reference customer willing to take a call.

The barrier moved from build to distribution. The last 8 weeks optimised hard for the side that stopped being scarce — 33,403 lines, 111 models, 77 pages, 0 customer conversations. That is not a criticism of the code, which is genuinely good. It is an observation about where the remaining risk sits.

---

✅ How to actually cash in the advantage

The low-cost-base insight is right. It just has to be spent on the correct constraint. Concretely, the advantage is real if and only if:

1. Never pay to acquire a customer. Ads are how the giants' losses are manufactured. 🎯 Business model decision of record: B2B-first — sell seats and engagements to companies, not courses to students #83 Phase B2 makes this a tripwire, not a preference.
2. Never salary a teacher or mentor. Revenue share only. Already built.
3. Sell few things at high prices. 30 sales/year at ₹1.64L is a business; 1,200 sales/year at ₹4,000 is a support queue with the same revenue and 40× the operations.
4. Sell to buyers who bring their own learners. An institution supplies demand and often supplies mentors — the only way a solo operator closes a two-sided cold start.
5. Automate the recurring ops before scaling volume, not after — payouts (Payout automation & creator KYC (RazorpayX) #14) is the first one that will hurt.
6. Shrink the maintenance surface. 34 unused models are a permanent tax on nothing.

Restated in one line: the advantage is not that the platform is cheap to build. It is that the business is cheap to run — and that only pays off if it is pointed at a buyer who does not have to be bought.

---

The honest bottom line

Nothing in the research says a solo operator cannot make good money here. Learnyst, Academically Global, Jaro and iMocha say the opposite.

What the research says is narrower and sharper: "build the platform and let supply and demand arrive" is the one strategy that has failed at every capital level from $5M to $159M. Not because those teams were bad, but because a marketplace's actual product is distribution, and distribution has never been something you can build. It has to be earned, one uncomfortable conversation at a time — and that is the part with no AI leverage and no shortcut.

The plan does not need to change. #83 and #84 are already built around exactly this constraint. What needs to change is the expectation that the hard part is behind us. The hard part hasn't started.