Leapfrogger.aiTop 10Deep-reviewed
by kishore bhatia
What they're building
B2C eCommerce agent: building a self-improving resale-items posting agent Your agent - haggles. One photo. Everything else is the agent’s problem. The agent does the boring part. All of it. It negotiates while you sleep. Self-improving AI - Every listing teaches the marketplace. Craigslist, rebuilt for agents.
AI code reviewrepo: real
Live deployment verified (bricklist-dev.up.railway.app returns HTTP 200, real 'Haggle' product, and the agent-readable GET /api/listings returns 200). All three sponsors are genuinely load-bearing in code: Tavily via @tavily/core client.search() in the enrich step; Nebius Qwen2.5-VL-72B through an OpenAI-compatible provider pointed at api.tokenfactory.nebius.com (identifies the set from a photo and voices every negotiation reply); and Composio via a hand-rolled v3 REST client (tools/execute to send Telegram replies + Svix webhook signature verification + connected-account owner resolution). The agent loop is sophisticated: a checkpointed identify→enrich→generate pipeline plus a code-owned, geometric-decay negotiation policy with floor guardrails, prompt-injection detection, episode/reward tracking, and settlement — covered by Vitest tests and ADR-backed docs. ⚑ Live eBay-publish and Telegram/Composio paths are wired but unverified end-to-end without keys (keyless mock fallback by design); judged on verified live web flow + code. Otherwise none material.