AndrewShatsky

Lokin

AI-native platform that connects college students into founder-ready teams. Scaled to 2,000+ users and 800+ weekly active users. Full-stack application with a modern web frontend and distributed backend services orchestrating profiles, startups, and team formation.

Problem

Match students into co-founder teams with no behavioral data and no ground truth for a “good fit.” Naive LLM matching was slow and sometimes invented reasons two people clicked.

Approach

An agentic layer over Postgres (profiles) and Pinecone (semantic similarity). An intent gate chose a full LLM call vs. cheap deterministic ranking, with a deterministic fallback so a bad model response never blocked the feed.

Tradeoff

Capped LLM usage over “always the best match” — slightly weaker tail ranking to hold a 2.0s p95 and predictable cost. Rolled out campus by campus to earn behavioral signal before trusting the model.

Outcome

2,000+ users, 800+ WAU through 25 campus organizations; the feed stayed responsive even when the model degraded. Honest miss: retention plateaued once novelty faded — recurring reasons to return, not match quality, were the real gap.

Fitness Tracker AI

AI wellness app and personal life coach. A computer vision pipeline logs calories and macros in 1.2 seconds from a photo, wearables sync automatically across 12+ integrations to track sleep and steps, and an AI agent cross-references health data to deliver fitness and lifestyle coaching. Macro estimates land within approximately 10% of manual entries in controlled testing.