I build AI-native products and software systems end to end — product scope, architecture, APIs, data and state, agent behavior, testing, and deployment.

My strongest work sits at the boundary between probabilistic models and deterministic software: structured outputs, tool use, retries, fallbacks, policy gates, idempotency, signed webhooks, evidence grounding, and failure-mode testing.

I use Claude Code, Codex, and Cursor as primary development tools for implementation, debugging, refactoring, repository exploration, and test generation. I still own the architecture and manually review the parts that cross trust, state, security, and execution boundaries before shipping.

Current focus: applied AI · agent systems · AI product engineering · fintech/payments · reliable LLM workflows

4x Hackathon Winner · BuidlGuidl Batch #23 · Built and deployed products across AI, fintech, EVM, Sui, Solana, Flare, XRPL, Mezo, Somnia, and Uniswap.

Deterministic execution firewall for AI coding agents.

Verifies AI-generated engineering plans against live GitHub repository state using strict AST static analysis, runtime preflight checks, and deterministic evidence rules before high-consequence code changes execute.

• Dual-Model Verification Architecture: Claude 3.5 Sonnet performs structural and semantic analysis while GPT-4o-mini runs bounded adversarial validation; deterministic software engines retain final authority over run pass/fail decisions
• Deterministic AST & Tree-Sitter Grounding: Cross-checks proposed diffs against repository ASTs, imports, exported call graphs, and file boundaries with zero model hallucinations allowed
• Async Distributed Pipeline: FastAPI orchestration with Dramatiq worker pools on Google Cloud Run, Redis Memorystore task queues, and MongoDB Atlas ACID-isolated run telemetry
• Enterprise-Grade Security: GitHub App signed webhook auth (RSA-SHA256), host-only strict cookies, token bucket rate limiting (10 runs/min), concurrent project run caps, and zero plain-text secret storage

AI-native negotiation checkout built around Razorpay.

A LangGraph negotiation agent handles buyer conversation and proposes commercial moves, while deterministic strategy and financial gates remain the only authority over price, agreement locking, and payment execution.

• Structured model decisions, bounded replanning, and deterministic safe fallbacks
• Server-side agreement revalidation and idempotent Razorpay Test Payment Link execution
• Raw-body HMAC webhook verification and canonical payment-state handling
• 93/93 backend tests covering negotiation, adversarial model output, prompt injection, concurrency, and payment/webhook safety

Proof-first job intelligence agent grounded in live role and candidate evidence.

Nuvra reads live openings from the Speedrun Talent Network, combines them with explicit public GitHub/portfolio evidence, and produces evidence-backed role matches, proof briefs, and application drafts.

• Bounded server-side search plan with deduplication and current role-detail retrieval
• Deterministic evidence-fit ranking; the LLM explains results but does not own the numeric score
• Zod-validated structured outputs, retries, grounding checks, and deterministic fallbacks
• Public GitHub/portfolio content treated as untrusted reference material with SSRF and CSRF protections

Cited company knowledge layer for marketing teams.

Turns brand documents into reusable structured knowledge cards, then uses retrieved evidence to answer questions and score/rewrite copy.

• Document ingestion and structured knowledge extraction
• Evidence-backed Q&A with source references
• Ogilvy-principle copy scoring and rewrite workflow
• Supabase/PostgreSQL state, authentication, row-level security, and usage/cost tracking

Non-custodial payment-versus-payment settlement for native XRP ↔ Flare assets.

A maker locks USD₮0 on Flare, the taker pays native XRP directly on XRPL, and Flare Data Connector proof is verified before escrow is released.

• Completed a real public-testnet XRPL → FDC → Coston2 settlement
• Dockerized Express API + async worker with persistent, restart-safe SQLite jobs
• Deduplication, rate limiting, payment preflight validation, replay protection, and recovery paths
• 59 automated checks across backend, contracts, and responsive browser behavior

Stablecoin-native receivables credit market on Arc.

Businesses sell buyer-accepted invoices for immediate USDC liquidity; lenders compete through funding auctions and escrow enforces the repayment waterfall.

Policy-protected AI intent execution on Sui.

Converts natural-language goals into typed execution plans, reads live DeepBook state, dry-runs transactions, scores risk, and enforces user-owned Move policy limits before execution.

Autonomous on-chain dispute resolution.

Research, validation, skeptic, and judge stages operate over public evidence with callback validation, retries, and escrow-backed settlement.

• SmartOnBoard — repository-aware developer onboarding with Python, Flask, IBM watsonx.ai, and GitHub analysis.
• RiskShield — Uniswap v4 impermanent-loss insurance hook built through the Atrium Academy × Uniswap Foundation Hook Incubator, Cohort 9.
• MezoPay — username-based MUSD payments on Mezo. 🏆 Winner, Encode Mezo 2026
• Aelys — AI copilot routing wallet/NFT analytics across bitsCrunch APIs.

Ambiguous product problem
        ↓
Define state, trust boundaries, and failure cases
        ↓
Use AI coding agents to implement and iterate quickly
        ↓
Review critical execution paths manually
        ↓
Add structured outputs / validation / retries / fallbacks
        ↓
Test adversarial and failure scenarios
        ↓
Deploy, exercise the real workflow, inspect what breaks

I prefer using the LLM for ambiguity, language, planning, and interpretation, while deterministic software owns permissions, state transitions, validation, and high-consequence execution.

Independent Applied AI & Full-Stack Developer · Remote · Jan 2024 – Present
Building and deploying self-directed AI and full-stack products across agent workflows, APIs, SQL-backed state, fintech/payments, and Web3 systems.

BuidlGuidl — Web3 Builder, Batch #23 · Remote · Dec 2025 – Jan 2026
Built an on-chain developer directory using The Graph and worked through deployment/debugging issues across Vercel and Arbitrum.