From 9 user interviews to a shipped product. An AI-native platform that diagnoses a PM archetype from a 10-question quiz, builds a personalised roadmap, and answers questions through a RAG chatbot grounded in a vector database I deployed and own end-to-end — not a managed RAG API.
This case study is the product write-up. The original Product Requirements Document — the “PM Career Navigation Platform for India” PRD submitted to Rethink AI’s MPM Cohort 7 (March 2026) — is the discovery foundation referenced throughout: 9 interviews, five user segments, and a problem-prioritisation chain. The Solution and Architecture sections below describe what was actually built and shipped.
PM Pathfinder began as a PRD for Rethink AI’s Managed PM Cohort 7 and became a real, live product. It is live at pm-pathfinder.gauravg.dev — quiz, archetype reveal, dashboard, roadmap, and a grounded RAG chatbot, all working.
This case study reads in two registers. Discovery is drawn from the submitted PRD — the research, segments, and problem chain that justified the bet. Solution and Architecture describe what I actually shipped: an archetype diagnostic, a personalised roadmap, and a retrieval-augmented chatbot grounded in a vector database I built and own. Where the live product and the original PRD diverge, this document says so plainly.
The audience is working professionals in India with 3+ years of experience trying to break into or advance in product management — engineers, consultants, narrow-domain PMs, designers, career-adjacent operators. Fresh graduates are explicitly excluded; the aspiration to become a PM forms only after real industry exposure.
The market runs from free (YouTube, WhatsApp groups, community Slacks) straight up to ₹2.5–4 lakh bootcamps. Nobody sells a monthly subscription of ₹2,000–5,000 that gives AI-powered practice, real feedback, and a believable readiness score. That gap is the entire opportunity.
Research basis: 9 interviews with 8 people, March 2026, plus community research across Reddit, LinkedIn, Quora, Grapevine, Glassdoor, and 20+ industry reports.
| Segment | Who They Are | What We Saw |
|---|---|---|
| Trapped Engineer | 2–4 yrs at IT firms, want PM, no one in their network knows it | 150–200 applications, 2–3 replies. Rates himself 15/100 while holding a product title. |
| Niche-Locked PM | Working PMs, 1–7 yrs, experience too narrow to transfer | Has the title but cannot show the job outside one domain. Largest group by likely revenue. |
| Methodical Planner | Consultants/analysts, 5–8 yrs, highly analytical | 2 years of active interest, zero applications. Knows every framework; cannot act without certainty. |
| Creative Crossover | Designers/marketers, 4–8 yrs, strong on users, weak on PM vocabulary | Made two study schedules, abandoned both. Applied to exactly one role, via a personal connection. |
| Perpetual Preparer | Adjacent to product 5–10+ yrs, knows the most, in the longest loops | 10 years of self-study, still calls himself “aspiring.” Deeply skeptical anything new will work. |
Research insight: the more someone knows, the longer they tend to stay stuck. Knowledge without a believable external signal of readiness produces preparation loops, not applications. The product’s first job is not to teach — it is to measure.
People are stuck because they cannot measure progress. Without a believable “you are ready,” the safest move is to keep preparing. Rejections arrive as a verdict with no explanation.
Everyone who moved into PM did it through a person, not an online application. Network access is real — but it’s a V2 problem. V1 fixes the root cause: practice and feedback.
A 4.5-year PM sitting through the same intro as a first-timer wastes both. The product must figure out who you are within 10 minutes and feel built for you.
All eight consumed huge amounts of content; none could perform. What moved the needle was practising with specific feedback. The centre must be a practice engine, not a library.
The only time anyone learns they aren’t ready is when they fail a real interview — the worst possible moment. Feedback has to be embedded into every practice session.
Turn studying into practice, give honest feedback, show a believable readiness score, and price it at ₹2,999/month — a point nobody else occupies.
The live product delivers the first half of the bet end to end — identity and personalisation — and lays the rails for the practice loop. A returning or new user moves through a single, value-first flow.
| # | Surface | What Happens | Design Principle |
|---|---|---|---|
| 1 | Onboarding + Diagnostic | 3-step context (background, years, industry) then 10 scenario MCQs across 4 skill dimensions. | Figure out who you are in <10 minutes |
| 2 | Results teaser | Blurred 5-dimension scores shown before sign-up. | Value before commitment |
| 3 | Archetype Reveal | One of 6 archetypes — Builder, Architect, Storyteller, Advocate, Operator, Strategist. | A specific mirror, not a generic grade |
| 4 | Dashboard | Multi-radar, stat cards, a readiness ring scored 0–100 vs APM / PM / Senior PM benchmarks. | Make standing legible at a glance |
| 5 | Roadmap + Topics | Gap-targeted learning path; 15 topic pages and per-dimension deep-dives. | The system picks what you work on |
| 6 | Navigator Chat | A RAG chatbot answers PM questions from real frameworks, filtered to your archetype. | Grounded answers, not guesses |
Where this diverges from the PRD, honestly: the submitted PRD centred a practice-problem engine and an AI feedback grader as the core loop. The shipped product nails identity, personalisation, and grounded Q&A first — and treats the graded practice engine as the next build (see Roadmap). The discovery thesis is unchanged; the sequencing reflects what was provable in a cohort timeline.
The Navigator chatbot is a real retrieve-then-generate pipeline, not a wrapper. Source
content — archetype playbooks and a “builder’s bible” of PM
frameworks — is chunked and embedded with Google’s gemini-embedding-001
into 3072-dimension vectors, then written to a document_chunks table in
Supabase Postgres with the pgvector extension. An IVFFlat index
(lists=50, cosine ops) keeps similarity search fast as the corpus grows.
At chat time the user’s message is embedded with the same model and matched against
document_chunks through a match_chunks() Postgres function —
a plpgsql RPC that ranks by cosine distance and optionally filters to the user’s
diagnosed archetype, always falling back to the shared “builder’s bible” so
general questions never come back empty. The top 4 chunks are stitched into the system
prompt before generation.
SDK removed entirely — the Node Gemini SDK hit Vercel serverless
runtime incompatibilities; hand-rolled fetch calls to the REST endpoints
removed that dependency.
SSE parsed by hand — the chat route reads
streamGenerateContent as a raw byte stream, splits on data:
lines, and re-streams plain text for a typing effect with no extra latency.
429 backoff built in — embedding and generation retry with exponential backoff (1s/2s/4s) before failing, since the free Gemini tier rate-limits aggressively under burst.
Service-role only, server-side — document_chunks
has RLS with no public read; only the server-side client (service-role key, never shipped
to the browser) can query it, and the route gates on an authenticated user before spending
an embedding call.
Auth is Google OAuth via Supabase. An anonymous quiz session is migrated to the authenticated user on sign-in through a single service-role route — the only place in the app that bypasses RLS.
| Table | Purpose |
|---|---|
| profiles | archetype, display name, avatar, Pro flag |
| assessments | dimension scores, archetype, tiers (JSONB) per user |
| quiz_sessions | anonymous pre-auth sessions + migration flag |
| learning_path_progress | step-level roadmap completion |
| deep_dive_results | per-dimension answers (JSONB) |
| portfolio_profiles / portfolio_case_studies | public portfolio snapshot, up to 3 case studies per user |
| document_chunks | RAG corpus — pgvector(3072), IVFFlat cosine index |
The free tier is the hook, not the product. Users get enough to feel understood and to see themselves improve before the paywall — which is framed around their specific gaps, not a generic upsell.
Note on current state: as a cohort case-study product, Pro-gating is intentionally disabled platform-wide — every feature is unlocked for all users while payments (Razorpay) are pending. The gating logic exists and works; it is simply not called yet.
Upraised charges ₹2.5L for human mentors. At ₹2,999/mo, even one human review per week per user erases the margin. The bet: AI feedback tuned to real PM interview standards closes 70–80% of the gap at ~5% of the cost.
Letting people choose what to study is the exact trap they’re in — all eight tried it and failed. The system always routes you to your weakest scored area. Users lose the feeling of control; they gain faster improvement.
Not a watered-down sample of the course — a real diagnosis. “You need 12 more sessions on metrics to be ready for mid-stage PM roles. Your plan is ready.”
Next.js App Router · TypeScript · Tailwind · Recharts radar
Supabase Auth (Google OAuth) · Postgres · Row-Level Security
pgvector · IVFFlat cosine index · plpgsql match_chunks()
gemini-embedding-001 (3072d) · gemini-3.1-flash-lite streaming · manual SSE
10-question scenario quiz · 6-archetype scoring · roadmap generation
Vercel serverless (60s max duration on the chat route)