If the first essay was the diagnosis — an application is a one-way message with no shared state — this one is the prescription. Fifteen ideas went into a RICE table; two bundles came out, and they are deliberately one loop rather than two features. The first changes where applications go. The second changes what happens to them. And what happens to them feeds back into where the next ones go.
Bundle A: apply where a reply is realistic
Every job card gets a Response Signal computed from behaviour Naukri already records: the recruiter's reply rate over the last seven days, per role; how recently they acted, shown in bands rather than timestamps; a Verified badge that means billed and KYC'd, so it cannot be bought by a scammer; "role live as of", derived from recruiter activity instead of the posting date; and "usually replies in two days". A stale role is not hidden — seekers told me they want to see it — but it gets an honest amber banner and three live alternatives.
The recruiter does nothing for any of this to appear. The only new recruiter behaviour in Bundle A is optional: reply to 30% or more of applicants within a week and earn a Responsive badge with a ranking lift. Silence is no longer punished; it is simply visible.
After applying, a small Apply-Smart sheet reflects: "You applied to seven roles this week; two had realistic odds." It never blocks, it appears at most once a week, and it offers one toggle to route future alerts toward realistic roles. This is how spray-applying falls without ever telling a fresher no.
Bundle B: see it move, confirm it both ways
Every application carries a six-state tracker — Sent, Seen, Shortlisted, Proposed, Confirmed, Attended — with an expected reply window from that recruiter's own history. "Seen" fires only when a recruiter opens the profile from the applicant list for at least five seconds; it never claims "read". A filled role closes honestly, with the reason and three live roles. Fourteen silent days archives the application and tells the seeker, so attention is freed.
| Recruiter tap | Where it lives | What the seeker sees |
|---|---|---|
| Propose | Row of the existing applicant list; a tiny sheet for three slots, mode, address | Pick a slot → Confirmed both sides → calendar + 24-hour and 2-hour reminders |
| Not now (bulk for fifty) | Same row, or select-all | An honest outcome instead of silence |
| Role filled | Same row | Closed with reason and three live alternatives |
| Mark interviewed | Day-of card | Paired with the seeker's "Yes, it happened" — the only thing that counts |
We store state, not judgement: no pipeline stages, scorecards, notes or interviewer assignment. Three verbs and one day-of confirmation. This is not an ATS, and it must never become one.
The handshake is where the metric is literally won. Confirmed appears only when the recruiter acknowledges the slot; Attended counts only when both sides say so. A seeker who cannot make it can propose another time or withdraw in one tap, and the recruiter's row updates. No pressure to book — a guardrail metric watches withdraw rates and complaint tickets for exactly that.
The AI piece: a parser, not a chatbot
Recruiter replies arrive as free text — "pls come for F2F Tuesday 4th Nov 11 am, Tower B Sector 62 Noida, ask for Priya". My vibe-coded artifact, the Interview Handshake Agent, turns that into a structured slot: date, time, mode, address, contact, with a confidence score. Confirm downloads a real calendar file, drafts a polite reply and sets the reminders. A fee-word shield runs first; "registration fee" or "UPI" turns the card red and disables confirmation. The parsed slot writes back to the timeline as "Proposed, seeker-reported", so even off-platform replies enter the two-sided flow. Deterministic rules come first because they are auditable, free and work on the Lite app; a structured-output model is the fallback for low-confidence messages, and it is forbidden from inventing a date that is not in the text.
The arithmetic
Per application, the counted-interview probability is response rate × proposal share × confirmation × attendance × recruiter marking. Today that is 11% × 65% × 62% × 75% × 82%, about 2.7%, which compounds over 8.6 applications to the brief's 21%. The levers are specific: response to 14.5% from the badge and one-tap actions; confirmation to 80% and attendance to 86% from the handshake and reminders; marking to 95% from the day-of card; applications down to 7.8 from Apply-Smart. The target case lands at 39%, the conservative case at 33%, and 36% sits between them. Every input is a yellow cell in the workbook, declared synthetic until Wave 0 instruments the real values.
On a 2.3-million-seeker monthly cohort that is roughly 0.35 million more counted interviews a month. For recruiters it means fewer no-shows and better-fit applicants, which is a renewal conversation account managers can have without a new SKU. Paid seeker products are held neutral by design — a guardrail reports the confirmed-interview gap between paying and free seekers at equal fit, and it must stay within a point.
Launch where the signal is dense
Wave 0 ships no UI: it instruments "Seen", recruiter actions and interview events and reproduces the 21% baseline. Wave 1 runs in Delhi-NCR and Bengaluru for Data, IT and Sales roles with verified recruiters who have at least twenty applications per role, on a 50/50 seeker holdout for eight weeks, with kill criteria of less than three points of lift or a rise in scam reports. Later waves add cities, role families, freshers, agencies, the Lite app and nine languages. The gates are metrics, not dates.
The brief ended with a line I kept on my desk: you can shape only one side of this market, and the other side owes the seeker nothing. My answer is to make the recruiter's existing behaviour visible, price the three actions that matter at one tap each, and count only what both sides confirm. A closed role and a withdrawn application are honest endings, not failures.
Prototype (nine routes, PM-annotation toggle, Handshake Agent): naukri-applied-to-interview-eswar.lovable.app. Deck, PRD, workbook and FigJam boards: case folder 14-naukri-capstone.