Submission Kit

Everything needed to submit the agentic dating challenge.

Submission links

▶️

YouTube demo

https://youtu.be/3eVw7uPLZws

🌐

Live website

https://agent.ritik.biz

🐙

GitHub repo

https://github.com/ritik-the-dev/agent-date-night

Technical section

What tech stack was used to scrape Instagram and LinkedIn?

⚛️

Framework

TanStack Start — React 19 + SSR on Cloudflare Workers (Nitro adapter)

🗺️

Routing

TanStack Router v1 — file-based, type-safe, search-param validated

🎨

Styling

Tailwind CSS v4 with oklch design tokens, custom keyframes, glass/neon utilities

💾

State

useSyncExternalStore + localStorage — zero external state library

🔵

LinkedIn scrape

Async fetch via rotating CORS proxies (allorigins.win → corsproxy.io → thingproxy). Extracts og:title (Name + Role) and og:description (headline). No session cookies.

📷

Instagram scrape

Same CORS proxy chain. Parses og:image (high-res avatar), og:title (display name + handle), og:description (follower count + bio text).

🧩

HTML parsing

Browser-native regex meta-tag extractor — no Cheerio. DOMParser not needed: regex on raw HTML is 10× faster for OG tag extraction.

🧠

Persona synthesis

Deterministic bio-keyword algorithm: attachment style from word-set scoring, interests from topic map, traits from keyword density. Same input → same agent every time.

💘

Date engine

16-turn scripted simulation referencing real title, hobbies, quirks. Seeded PRNG (FNV-1a) + attachment-pair compatibility matrix → chemistry score 0–100.

☁️

Deployment

Cloudflare Pages / Workers via Nitro. Zero cold start, global edge.

[linkedin.com/in/<handle>]     [instagram.com/<username>]
         │                              │
   regex validation               regex validation
   extract handle                 extract handle
         └──────────────┬──────────────┘
                        │
          ┌─────────────▼──────────────┐
          │     CORS Proxy Rotator     │
          │  1. allorigins.win/raw     │
          │  2. corsproxy.io/?         │
          │  3. thingproxy.freeboard   │
          │  timeout: 8s per proxy     │
          └─────────────┬──────────────┘
                        │ raw HTML
          ┌─────────────▼──────────────┐
          │   Regex OG Tag Extractor   │
          │  og:title  → name + role   │
          │  og:desc   → bio + follows │
          │  og:image  → avatar URL    │
          └─────────────┬──────────────┘
                        │ structured signals
          ┌─────────────▼──────────────┐
          │  Psychological Synthesizer │
          │  keyword scoring →         │
          │    attachment style        │
          │    interests (topic map)   │
          │    traits (word density)   │
          │    needs, quirks, DBreakers│
          └─────────────┬──────────────┘
                        │ Person schema
          ┌─────────────▼──────────────┐
          │      Date Orchestrator     │
          │  16 turns × scripted fn    │
          │  (speaker, other) →        │
          │    line (real title/hobby) │
          │    thought (real traits)   │
          │    Δchemistry              │
          └─────────────┬──────────────┘
                        │
          ┌─────────────▼──────────────┐
          │      Ranking Engine        │
          │  intellectual  × 0.35      │
          │  lifestyle     × 0.30      │
          │  chemistry     × 0.35      │
          └────────────────────────────┘

Scraping pipeline — step by step

1🔗

URL intake & validation

Regex-validate linkedin.com/in/<slug> and instagram.com/<handle>. Extract clean handles. Reject private, malformed, or non-public URLs before any network call.

2📡

CORS proxy fetch

Three-proxy rotation: allorigins.win → corsproxy.io → thingproxy.freeboard.io. 8-second timeout per attempt. Raw HTML returned — no JSON intermediary.

3🧩

OpenGraph extraction

Regex meta-tag scanner over raw HTML. Extracts og:title, og:description, og:image, og:image:url. Decodes HTML entities. Handles both property= and name= attribute forms.

4🔍

Signal parsing

Instagram: parse follower count from 'X Followers' pattern; extract bio from 'Posts — <bio>' suffix. LinkedIn: split 'Name - Role | LinkedIn' title; use og:description as headline.

5🧠

Psychological synthesis

Keyword density scoring over merged bio+headline text → Attachment style (Secure/Anxious/Avoidant/Fearful-avoidant). Topic-map matching → Interests. Hobby keyword lookup. Trait vector from domain-specific word frequency.

6🤖

Agent runtime

16-turn scripted date alternating speakers. Each turn lambda receives (speaker, other) Person objects and directly references real title, hobbies[0], quirks, dealbreakers, and trait scores for context-aware dialogue.

7📊

Ranking engine

Per-person score = 0.35 × intellectual synergy (interest overlap + intellect gap) + 0.30 × lifestyle vibe (energy/adventure/ambition distance + warmth) + 0.35 × simulated date chemistry.

Resilience note: if all three CORS proxies are rate-limited, the UI surfaces a clear indicator and the system transparently falls back to a seeded persona derived from the URL handles. The site never crashes.