Your dating profile isn’t failing because you’re uninteresting. It’s failing because you’re playing a visibility game with hidden rules while everyone who’s winning is quietly running experiments.
Zesty doesn’t guess. It extracts signal from noise and returns a calibrated score with a confidence you can read like a weather report. Then it gives you one high‑ROI move that actually changes outcomes.
The Architecture of Clarity (not vibes—math)
Traditional “AI” products shotgun prompts at a dozen models and pray. We do the opposite: a lean, deterministic engine that uses small, surgical LLM calls only where human language matters. Three stages, zero fluff:
-
Stage 1 — Data Separation:
- Calculated metrics: VCI, confidence (0–1 and 0–100), effective sample size, rankings, cohort position, score trajectory.
- LLM inputs: votes joined to voter segments, technical photo features, market sample, user voting behavior.
- Context: profile facts, declared vibes (50/30/20 weights), preferences, test metadata.
-
Stage 2 — Deterministic Analysis: Ordinal‑Bayesian scoring with vibe priors, reliability weighting, MRP correction, calibrated uncertainty, photo vs profile attribution, polarization, and an ROI engine that ranks actions by predicted lift.
-
Stage 3 — Narrative Synthesis: Two small photo analyses (UI + technical) feed a final, evidence‑gated master prompt that turns the math into a premium report you’ll actually read and use.
The Four‑Level Decision That Mirrors Real Swipes
Star ratings are fake precision. We use the decision people actually make: Pass / Unsure / Maybe / Yes. That maps to a latent signal η with a cumulative‑link (ordinal logit) model trained on historical data.
Minimal math, maximum sanity:
P(Y ≤ k | η) = σ(c_k − η) with c_1 < c_2 < c_3
E[VCI] = Σ_k P(Y = k) × anchor_k, where anchors = [1, 4, 7, 10]
Result: a clean 1–10 score that isn’t a vibe check; it’s a calibrated estimate with error bars.
Weighted Voting: Trust the Signal, Not the Loudest Clicker
- Time decay: recent votes weigh more (90‑day half‑life). Attraction drifts; math keeps up.
- Voter reliability: historical calibration → reliability weights (clipped; no oracle voters).
- Effective Sample Size (ESS):
- The honest truth about how many independent minds your score represents.
- High ESS = narrow range and higher confidence. Low ESS = bigger range, proceed with humility.
MRP Correction: Your Score, But Market‑Representative
If this week’s voters skew 18–24 and you’re aiming 28–35, we reweight segment posteriors to your target distribution. You see raw vs MRP‑adjusted scores and the bias magnitude. Political pollsters call this table stakes. Dating apps should, too.
Confidence, Translated to Human
We expose uncertainty the way sane systems do: as a percentage tied to ESS and posterior variance. A tight 95% range with 80%+ confidence means “stable, repeatable.” A wide range with low confidence means “promising, but get more quality votes.”
Read it like a grown‑up:
- 7.2 (95%: 7.0–7.4) at high confidence → ship changes with conviction.
- 7.2 (95%: 6.3–8.1) at low confidence → gather more signal before declaring victory.
Vibes (Structure, Not Stereotypes)
Users select up to three archetypes—Adventurer, Athlete, Builder, Charmer, Connector, Creative, Cultural Bridge, Intellectual, Minimalist, Tech Innovator—weighted 50/30/20. Each carries a learned offset (and variance) that forms a prior for your profile’s latent signal.
- No gender in priors. Your baseline comes from personality + portfolio, not chromosomes.
- Why it matters: vibes stabilize early scores and let us learn faster from fewer votes.
Photo vs Profile: Who’s Carrying the Team?
Votes are nominally for the whole profile, but humans anchor on the first photo. We estimate an attribution coefficient λ:
η_vote = λ × θ_photo + (1 − λ) × θ_profile
You’ll see lines like “Main photo: 62% of appeal.” Translation: if λ is high and your primary is weak, changing one image can move your whole score.
Polarization vs Safe Score
We compute entropy of the vote distribution to quantify risk.
- High safe score → aligned brand, consistent reactions, easier to scale.
- Low safe score → split audience (edgier profile), risky but powerful when targeted.
What the Engine Actually Ships
You don’t get a 20‑page PDF full of lorem ipsum. You get an executive verdict, a ranked action plan, and photo portfolio triage—all tied to evidence.
- Emergency swaps: delete now / make primary (with predicted uplift).
- Winners & liabilities: per‑photo keep/optimize/delete with strengths/issues.
- Market: your cohort rank, gap to top performer, and whether you’re stuck in a crowded band where small changes cause big rank jumps.
- Mindset patterns: flags for bio sabotage (defeatist tone, crude content, passive framing), translated into practical rewrites.
- Conversation toolkit: openers that hook into your photo contexts so you stop fumbling after the match.
Built Different (Engineering Notes)
- Deterministic first: eleven rule‑based factories compute the truths (priors, VCI, ESS, MRP, attribution, polarization, winners/liabilities, ROI).
- LLM, but tiny: two async photo passes (UI + technical), then one master synthesis constrained by evidence—no raw vote quoting, no hallucinated numbers.
- Calibrated forever: we track CI coverage, confidence monotonicity, and action utility (expected vs observed lift). When drift happens, we recalibrate confidence curves—not the math.
The One Move That Usually Moves the Needle
Most profiles don’t need a personality transplant. They need a photo order correction plus a bio de‑sabotage.
- Swap your true winner to primary (we’ll tell you which). Typical lift: +0.8 to +1.5 VCI.
- Rewrite the first 150 characters: remove pre‑rejection and crude lines; lead with one specific, authentic detail that creates a conversation hook. Typical lift: +0.3 to +1.0 VCI.
Ten seconds of work. Real, measurable change.
Read Your Confidence Like a Pro (Cheat Sheet)
- Confidence 80–95% + tight band → Implement the top action now; revisit in a week.
- Confidence 55–75% + medium band → Implement, then gather 20–40 more quality votes to confirm.
- Confidence <55% + wide band → Don’t overfit. Collect signal; avoid sweeping changes.
Why This Works When “Advice” Doesn’t
Because it treats dating like what it is:
- An information problem (hence profiles)
- A signaling game (hence photo selection)
- A matching market (hence MRP and cohort ranks)
- A narrative exercise (hence a small, surgical LLM to translate math to motivation)
Zero Hallucinations, Maximum Receipts
Every spicy line in the premium report is chained to evidence: drivers, badges, confidence, and pre‑computed actions. If data is thin, the report literally says “insufficient evidence.” That’s not a bug. That’s integrity.
Your Roadmap (Four Weeks to Noticeable Change)
Week 1: Swap primary, delete liabilities, rewrite bio opener → +25–40% matches.
Weeks 2–3: Shoot one professional and one social proof context → trust jumps; conversations get warmer.
Week 4: Tune for your market (MRP shows where you win) → better dates, fewer dead chats.
TL;DR
Your profile is not a personality test. It’s a signal processor in a noisy market. Zesty’s ordinal‑Bayesian, reliability‑weighted, MRP‑corrected engine pulls the truth out of chaos and tells you the one move that actually matters.
Stop guessing. Start knowing.