Full-stack software engineer building live products across web, mobile, backend systems, real-time infrastructure, SDKs, developer tools, AI-agent workflows, and AI evaluation systems.

I built and launched Agentic and SolPulse: two production products across web, iOS, and Android with 2,250+ users, 62 reviews, 4.2★–4.3★ ratings, and live App Store, Google Play, and Solana dApp Store releases.

AI-agent wallet application where agents prepare wallet actions and external payment requests while users approve through their own wallets. Live across web, iOS, Android, desktop, CLI, MCP tools, and SDK surfaces.

• Traction — 1,250+ users in 2 months, 4.2★ rating, 36 reviews.
• Non-custodial approval model — Agents propose actions; users approve with their own wallets; private keys never leave the device.
• Agent workflows — Research tokens/wallets, check market conditions, set constraints, and prepare sends, swaps, limit orders, DCA schedules, lending/borrowing moves, and message/transaction signing.
• Shared agent/payment architecture — One WalletBackend powers browser app, mobile paths, desktop shell, CLI, MCP tools, Vercel AI tools, Solana Agent Kit adapter, A2A AgentCard metadata, and AP2/ACP/MPP payment- adapter flows.
• Production signing UX — Approval polling, transaction simulation, sign-and-send flows, Wallet Standard support, and live wallet review cards inside chat.

Solo-built production trading platform across web, iOS, and Android with 1,000+ users, 26 reviews, and a 4.3★ rating.

• Execution engine — DCA, limit orders, smart TP/SL exits, routing, priority-fee controls, execution logs, and 1,200 concurrent transaction monitors at 10ms intervals.
• 35 intelligence signals — Token risk, liquidity, smart-money activity, dev identity, rug/fraud detection, whale activity, and real-time filters.
• Risk controls — Max daily loss limits, per-trade risk caps, portfolio buckets, and regime-aware controls exposed through a consumer UI.
• Full stack — React/Vite, Node/Express, PostgreSQL/Prisma, Redis, WebSockets, SSE, backend APIs, database design, frontend UX, mobile releases, and production support.
• Non-custodial signing — Wallet approval flows where private keys remain inside user wallets/devices.

Model-to-capability runtime for legacy back-office applications: an LLM drives a workflow once, the run is converted into a typed/versioned capability artifact, and future executions replay deterministically with no model in the decision path.

• Model-free replay — Verified a live Claude Opus discovery run whose synthesized artifact replayed without a model, separating ok, typed business outcomes, recoverable states, and hard failures.
• Legacy UI surfaces — Built browser and terminal/green-screen surfaces with normalized observations, typed actions, policy chokepoints, evidence journals, redaction canaries, and CLI replay.
• Conformance with teeth — 25 hostile fixture scenarios killed 9/9 weakened replay engines, including 13 false-success failures where a broken engine would have reported success for an unsafe or incomplete run.

6/6 frontier agents failed with clean reward-0 results. Original Terminal-Bench 3 task modeling a durable approval outbox: turning an approved AI action into an executed external action exactly once while workers crash mid-call, deliveries duplicate, approvals are revoked, and the external tool returns UNKNOWN before an authoritative receipt arrives.

• 6/6 clean reward-0 trials — 3 Claude Opus runs and 3 OpenAI Codex runs failed without timeouts, API errors, or orphaned runs; the reference implementation passes 267/267 checks.
• Real false-positive failures — Agents produced locally plausible implementations that still mishandled UNKNOWN outcomes, receipt reconciliation, revocation, exactly-once execution, progress, or audit-state legality.
• Adversarial verifier hardening — Built isolated tool-ledger collection, process separation, root-only controls, unprivileged execution, cheat oracles, and deterministic reward checks to prevent reward-hacking paths.

Open-source systems for durable external actions, runtime crash behavior, benchmark verifier integrity, and success-judge reliability.

• Durable Agent Outbox — TypeScript infrastructure for AI agents that execute real external actions without duplicating side effects during crashes, retries, timeouts, UNKNOWN outcomes, revocation, or receipt reconciliation.
• Crashpoint — Crash/recovery conformance kit measuring exactly-once external side effects across LangGraph, Temporal, and DBOS through an out-of-process ledger the runtime cannot forge.
• Cheat-Oracle — Defensive verifier-hardening audit showing how in-container answer-key detectors can undercount benchmark reward hacking, plus a tamper-evident out-of-container detector design.
• Judge-Artifact — Eval-integrity audit measuring how shipped success graders over- and under-credit identical agent transcripts, including real prompt-injection ASR inflation on AgentDojo data.

Invited by AI Fund as one of 12 engineers to build against a B2B sales problem posed by a Fortune 100 OEM.

ConfigPilot is a decision engine for commercial configuration and pricing. An LLM proposes recommendations; a deterministic layer verifies each one against margin floors, live inventory, lead-time feasibility, and policy rules before it reaches a seller. Anything failing a check is blocked or flagged with an audit trail naming the constraint that caught it.

• The model never has final say - every recommendation passes through business rules that can veto it, so an LLM error becomes a blocked suggestion instead of a bad quote.
• Full-stack delivery - frontend, backend, validation logic, recommendation review, and approval flow, shipped as a working demo with a PRD and technical design doc.
• Why it's built this way - a wrong OEM quote costs margin or breaks a delivery commitment, so correctness has to be enforced outside the model rather than prompted into it.

• Agent Context Containment — TypeScript policy engine that blocks indirect prompt-injection attacks by tracking source provenance, tool capability, argument role, and approval receipts.
• Model Regression Sentinel — TypeScript CLI for detecting silent model-alias behavior drift by rerunning frozen prompt corpora against baselines and confirmation runs.
• Toolcall Risk Classifier — Python risk-scoring and evaluation harness for proposed AI-agent tool calls, with conformance tests, leakage audits, and benchmark-style failure analysis.

Swift Package for iOS-native wallet deeplink signing across Phantom, Solflare, and Backpack.

• Encrypted transport — X25519 ephemeral keys, NaCl box encryption, Base58 encoding, callback decoding, nonce/session handling, and wallet error mapping.
• iOS lifecycle support — SwiftUI/UIKit URL routing, pending request cancellation, Keychain-backed state, and deterministic smoke-test logging.
• Wallet picker UX — Phantom, Solflare, Backpack, and Jupiter Mobile support with remembered wallet preference.
• MWA-style app semantics — authorize, deauthorize, signMessages, signTransactions, signAndSendTransactions, getCapabilities, and SIWS-style flows over iOS deeplink transport.

Open, embeddable pre-sign transaction safety system for wallets, dApps, trading UIs, mobile apps, and agent signing flows.

• Static analyzer — Decodes raw transactions locally and flags risky patterns before signature.
• Drain detection rules — Durable nonce abuse, authority changes, token approvals, system assigns, suspicious writable accounts, and wallet-drainer primitives.
• Simulation and diff layer — Optional RPC enrichment for inner-CPI walking, balance/owner/delegate diffs, and divergence checks.
• Embeddable packages — TypeScript core, rules, registry, simulation, React hooks/components, and CLI scanner for browser apps, Node, React Native, Capacitor, and CI gates.
• Proven on mainnet - atomic-guard behavior verified against real mainnet transactions, not just fixtures.
• Test suite - [N] tests covering rule detection, simulation diffing, and the embeddable packages.

First Mobile Wallet Adapter SDK for Cocos Creator, built from scratch.

• TypeScript API, native Android bridge, ECDH/AES-GCM wallet association, SIWS, message signing, transaction signing, sign-and-send flows, auth caching, and multi-wallet support.
• Hardware-verified with Phantom, Solflare, Backpack, Jupiter, and Seed Vault on Seeker.
• Includes Token Duel demo game proving real on-chain gameplay flows.

First production-ready Mobile Wallet Adapter plugin for Unreal Engine 5.

• Kotlin clientlib, C++ JNI bridge, Blueprint UFUNCTION surface, auth caching, SIWS, signing, sign-and-send, capabilities, and multi-wallet support.
• Exposes native MWA, sign-then-broadcast fallback, and auto routing for wallet compatibility.
• Hardware-verified with Phantom, Solflare, Backpack, Jupiter, and Seed Vault on Seeker.

Capacitor plugin for native Android Mobile Wallet Adapter support from hybrid mobile apps.

Multi-wallet Godot example app and SDK integration work.

• Seeker hardware verification across Seed Vault, Phantom, Solflare, Backpack, and Jupiter.
• Three merged upstream PRs for reconnect/full state reset, getCapabilities, sign-and-send bridge support, SIWS wallet-proof APIs, and auth-token caching.
• Found 10 integration bugs during hardware testing.

Unity wallet adapter example app and SDK integration work.

• Found 11 integration bugs during hardware testing.
• Built auth-token caching, reconnect flows, deterministic logging, and MWA 2.0 API coverage.
• Merged upstream WebGL wallet auto-select UX fix.

TypeScript, JavaScript, React, Vite, Node.js, Express, PostgreSQL, Prisma, Redis, WebSockets, SSE, REST APIs, Swift, SwiftUI, Kotlin, Java, C#, C++, GDScript, Capacitor, Unity, Unreal Engine,Godot, Cocos Creator, SDK engineering, CLI tools, MCP, Vercel AI SDK, Playwright, Docker, conformance suites, deterministic interpreters, policy gates, wallet integrations, non-custodial signing flows, and production support.

• Portfolio
• LinkedIn
• Twitter / X