Case Studies

An MCP tool description is the one piece of text a model reads on every single call, which makes it where you stop an agent misreporting your numbers. I shipped one for AddressIntel's permit data, then found two of my own instructions to the model had already gone stale.

A swap's calldata often carries a few bytes past the last ABI word boundary, and contract decoders drop them. Eight of fifty live Ethereum swaps carried such a tail. All eight were booked fingerprinted and unattributed, and that is the design.

A rejected portal login returns an empty permit list and the run reports success, so a broken scraper and a quiet city produce the same zero. A silent result-set cap hands back the oldest 500 rows and drops the newest. Here is the permit pipeline built to detect both, across 20 permitting jurisdictions and six vendor portal platforms.

A week tuning a classifier's precision while two thirds of the target volume sat outside a stale candidate window. Three plausible heuristics that all failed, one narrow rule that survived measurement, and four of my own wrong results that shared a shape.

Three quiet failure modes in analytics platforms: the cost allocation whose residual is zero by construction, the identity join that merges two customers, and the fact table that changes months. Each one runs green.

Neutral, third-party proof of where DEX volume really originates, surfacing hidden, untagged orderflow plus execution quality and MEV across Ethereum, Base, Arbitrum, and Optimism. Live API + dashboard.

Predictive real-estate intelligence fusing fragmented municipal permit records with multimodal AI (Google Gemini) to score teardown probability, flippability, and off-market opportunity. Live B2B API.

A real-time anomaly-detection oracle that scores the reliability of decentralized GPU compute (Akash, Render, io.net) as a 0–100 Compute Health Score for autonomous AI agents.

A FastAPI rendering engine that aggregates Notion databases and streams dynamic, real-time SVG charts directly into Notion embeds.

Thirteen contracts on Base forged five of Coinbase's tokenized-equity tickers for $393,805. The address prefix, the token name and the bytecode all fail to separate them from the genuine ones. Reading the backing works, and watching it move works better.

1,005 tokens on Ethereum mainnet sharing one bytecode and two deployer wallets, 184 of them impersonating a canonical ticker across 143 distinct real tokens, and how a Uniswap V3 initialization tick turns 0.000739 WETH of real assets into a $3.28 billion liquidity listing that survived six months on DexScreener.

A practical methodology for detecting MEV and sandwich attacks on-chain, separating real slippage from adversarial extraction, across Ethereum, Base, Arbitrum, and Optimism.

How AddressIntel unifies building permits from four municipal systems on the parcel APN, engineers a teardown label that is independent of the feature it validates, and backtests it point-in-time, with an honest read on what the model actually beats.

Nine structural bets about execution, MEV, and data verification on-chain that don't depend on any specific protocol, model, or chain staying the same.

How ClearTrace turns anonymous, high-volume "Unknown Proxy (0x…)" contracts into named entities with a cheapest-first resolution ladder across a dozen sources: on-chain registries, bytecode fingerprints, proxy slots, and paid APIs tried last.

Off-chain RFQ and private-market-maker fills revert under fork-simulation. ClearTrace detects them from the quote's own route and sends them to an on-chain read instead of discarding the sample, so the execution-quality leaderboard is not biased toward AMM-only flow.

A revert rate is a reliability signal only where the failing transaction is the user's own, so ranking it across execution models flatters the venues whose failures happen out of sight. ClearTrace encodes an execution-model taxonomy that refuses to rank routers, batch auctions, and intent venues against each other.

Most predictive-ML backtests quietly leak the future and report a number that falls apart in production. Here is the point-in-time discipline that separates a defensible model from a flattering demo: the as-of-T rule, time-based folds, rare-event metrics, label independence, and a pre-registered go/no-go.

A blockchain has no sessions and no user IDs, only a flat log of swap legs. Window functions rebuild each wallet's ordered path between DEX entrypoints, but the work that decides whether the output means anything happens before the first LAG: collapsing legs to one transaction, excluding arbitrage bots, and keeping each chain's timeline separate.

There is no fee event on a swap. A frontend's cut is an ordinary ERC-20 transfer hiding among the swap's own token movements, so identifying it is a problem of exclusion rather than decoding: five categories to rule out, an asymmetry that makes a fee sink a signal in one direction only, and a confidence ceiling that stays at medium on purpose.