I build the systems that make products scale

Built a fundamental analysis Telegram bot for end-to-end trading on Polymarket, wiring real-time data feeds, prediction market APIs, and automated execution for event-driven bets.

Analysed trading strategies across historical datasets, surfacing patterns in entry/exit signals, win rates, and risk metrics, and integrated that analysis into the live product.

Designed and built the proof generation pipeline from scratch, end to end.

Developed the core contracts driving lending pools, credit scoring, and loan lifecycle management.

Built backend services in Go and Fastify for large-scale data processing and system orchestration.

Shipped the Next.js dashboards borrowers and lenders used to operate the product.

The iOS app joins a private tailnet through an embedded tsnet module, registering with the key from a QR scan; the tunnel terminates at 127.0.0.1, which also sidesteps iOS ATS.

Self-hosted Headscale control plane running database policy mode with per-account ACL isolation. The phone and the Mac's sakura-tsnet sidecar both register there and meet over the WireGuard mesh.

Rust API server handles auth, GitHub, Convex session reads, and the CLI device flow, and auto-provisions each account's tailnet key when the daemon starts.

Node daemon reads ~/.claude, ~/.codex, and ~/.local/share/opencode, and drives the claude, codex, and opencode CLIs locally.

Caddy fronts both services with auto-HTTPS; the Next.js web app on Vercel handles CLI login and invites.

A parent orchestrator routes each message to exactly one specialized sub-agent with a whitelisted tool set. No sub-agent can dispatch, so the tree is exactly one level deep and there is no fan-out.

Memory is bucketed by decay speed across seven semantic segments; effective importance decays exponentially from last access, and the nightly pass prunes anything under 0.08. Accessing a memory resets the clock and can promote its bucket.

A 03:00 launchd job runs a three-stage adversarial pass over the store: a consolidator proposes keep/promote/merge/prune, an adversary challenges each proposal, and a judge rules on the contested ones, biased toward preserving identity and relationship segments.

Every action writes an immutable event and a record snapshot, so the whole history is reversible.

Captures journal entries, schedules reminders, watches X handles with free-form filters on a 5-hour digest, serves weather from Open-Meteo, and reads Gmail behind a redaction policy that strips OTPs and financial numbers before the LLM ever sees them.

Each run is bounded by a per-sub-agent USD spend cap and turn cap. When “ok, logged” is the whole response, it sends a tapback instead of typing.

Built to solve my own problem of evaluating founders before joining their startups.

Scores founders based on project stage, profile quality, grants, and tweet authenticity.

Analyzes key follows including ecosystems, investors, and VCs from an X profile.

Helps freelancers and builders quickly assess if a founder or project is worth betting on.