About
I build data-intensive applications โ systems that take messy data and make it legible โ including AI: agentic harnesses where the model reasons and a deterministic engine computes.
The same rigor runs through my data work. I've traded crypto since 2020, and when I had to do my taxes, I realised nothing could reconcile my transactions across 20+ wallets and multiple chains โ so I built my own tools. One is a trade planning, execution, and analysis engine: a deterministic double-entry ledger that books crypto and fiat together, derived from rules and raw events and regenerated rather than hand-edited, so every number carries its receipts โ FIFO cost basis, per-wallet provenance, audit-grade by construction.
When FTX collapsed in 2022, I realised I was holding a rug recovery tool: the proof lived off-chain โ funds frozen on an exchange that never touched a wallet โ but the same ledger could book that loss, structured and sourced. That set me experimenting with early local LLMs and ML: a multi-modal pipeline (voice transcription, vision, and OCR over screen recordings, PDFs, and statements, all on local models) that turns that unstructured evidence into structured, sourced entries.
When AI matured with Claude 3 in 2024, I put the model in the loop โ reasoning over messy, unstructured evidence, reconciling it across sources, and classifying it into the ledger's schema, while the deterministic engine keeps the books. The same engine reconstructs any position from raw on-chain events for analytics and due diligence, every figure traceable to source.
Around it sit Solana DeFi P&L analytics, on-chain event tooling, and Xero financial reporting. Iโm a hardware and software power user, and I share what I learn about AI and tooling for staying productive solo.
Open to software-engineering roles in AI and crypto โ remote, or based in Tel Aviv.
Work
Self-employed
Tech Stack
Gallery


The engine
The private engine behind the tools here โ a deterministic double-entry system for crypto and fiat where the ledger is derived from rules and raw events and regenerated rather than hand-edited, so every balance is reproducible and audit-grade. It's the deterministic half of the AI harness in my bio โ the model reasons over messy evidence, this engine computes the books. FIFO cost basis, multi-chain event parsing, per-wallet provenance, multi-entity consolidation, price resolution, Australian CGT, and Xero integration on the fiat side. Its reconstruction core powers the per-collapse loss-quantifier sites below. Built because no existing tool could handle DeFi positions, collapsed exchanges, and SMSF compliance.
The harness
The model never touches the books. It proposes inputs โ classified transactions, matched sources, entries extracted from messy evidence โ and the deterministic engine regenerates the ledger from them: double-entry must balance, every figure traces to an event and a rule, and a wrong proposal is deleted and rebuilt, not patched (Claude Code plus domain skills). That is how non-deterministic AI reasoning coexists with a ledger that has to survive a tax audit. The same harness pattern, pointed at web development instead of accounting, built this site; the build log tells that story. Interactive demo coming: recorded agent traces replayed through a live in-browser eval scorer โ no live LLM calls, by design.


A Meteora DLMM liquidity simulator: choose a price range of discrete bins, a distribution shape (spot / curve / bid-ask), and which side to fund, then scrub a seeded price walk. A generator derives per-bin liquidity, fee accrual, and impermanent loss vs HODL โ regenerated from rules, never hand-edited, and reproducible from the seed.


A delta-neutral strategy simulator: feed strategy params and market data, and a generator derives the full trade ledger plus a scrubbable PnL timeline. Generator-first โ the ledger is regenerated from rules, never hand-edited, and stays delta-neutral by construction.


A turn-based on-chain game where a bot plays automatically โ each match emits a win/loss event, and the rewards/PnL ledger is derived from those events plus the payout rules (stake, payout multiple, protocol rake), regenerated rather than hand-edited. Generator-first, like the trade ledgers: every balance traces to a match outcome and a rule. Off-chain MVP now; same generator behind a Solana testnet program + event listener next.


The edge cases that silently break a normal crypto ledger โ internal transfers between your own wallets, chain reorgs โ reconcile correctly because the ledger is generated from raw events, not hand-posted. Minimal single-concept demos: phantom taxable gains vanish when self-transfers carry cost basis, and a hand-posted ledger diverges from a derived one the moment a confirmed tx is orphaned.
Tool ยท AccountMD engine
A hub of per-collapse loss quantifiers (FTX, Terra/LUNA, Celsius, โฆ): a victim pastes a read-only public wallet address and the AccountMD engine reconstructs their history into a quantified AUD loss, then a CGT-ready evidence report โ every figure sourced to an on-chain transaction. Read-only, address-only: no wallet connection, no signing. Programmatic-SEO, one landing page per collapse (Marc Lou playbook).
Building in Public
I write about building crypto and financial-data tools โ and what I learn using AI to ship more on my own.