Solomon exhausted 95% of a Claude Max session allowance. The agents were not repeating identical work. They were keeping long conversations alive across hundreds of tool calls, so Claude repeatedly processed the growing history. This post measures the waste and proposes bounded sessions with compact handoffs.
In 2022 I swapped the cryptographic library under DynamoDB’s TLS termination from OpenSSL 1.0.2 to AWS-LC and deployed it to every AWS region with zero downtime, on a database serving hundreds of billions of requests an hour. Median handshake time dropped 12–35% by region, and FIPS endpoints dropped as much as 80%. This is the internal ops-win announcement rewritten for the public, including what the integration required — modifying Nginx source against a new crypto API, a legacy build system, Amazon’s deployment machinery, and cryptographic correctness requirements — and the two problems the rollout surfaced.
bitcoind can publish new blocks and transactions over ZMQ, and tools like LND depend on it. But if a configured ZMQ socket can’t bind at startup, bitcoind starts anyway and only logs it at debug level — so the problem surfaces days later, in someone else’s process. This is the bug, the 2019 history behind it, and the two-function design that fixes it.
At Amazon I designed, built, tested, and deployed public phone-number APIs for Amazon Connect, in Java, released globally, for Connect users provisioning numbers for their contact centers. In Connect a phone number is a cloud resource with an ARN, tags, an asynchronous provisioning workflow, and a quota rule limiting how often numbers can be claimed and released. This post covers the API surface and the design decisions encoded in the contract.
A Hidden Markov Model assumes the market is always in one of a few states you cannot see directly. I fitted one to order-flow data to look for states that lasted long enough to trade. Given one input it called direction correctly 50% of the time, the base rate. Given several inputs, the states it found were separated by how much the price was moving and not by which way it was moving, and the small directional signal in the raw input did not survive. This post covers what the model does, how to run inference without using future data, and what the added complexity cost.
I implemented a 2010 order-flow paper on four years of tick data and reproduced its headline result: one number explains about two-thirds of contemporaneous price movement, an R² of 67%. I then tested six strategies for trading that signal. All of them were negative after the bid-ask spread, across four years of data.
I built a soccer Elo model, published it as an in-browser predictor, and spent most of the project testing whether it could beat the bookmaker’s closing line. It does not. Over 4,099 club bets the closing line value was −3.4% with the entire confidence interval below zero, which measures how efficient these markets are.
A trading bot is a distributed system that runs unattended against a live broker, so a crash, a duplicated message, or stale local position state each has a direct financial cost. This is how I engineered one to run all day under five reliability guarantees: idempotency, a single source of truth, bounded-cost orders, clock-aware failover, and self-monitoring that catches a silent break within minutes.
Twenty-five Pine strategy files and dozens of indicators. Most measured a real pattern, but none of the patterns held across a different test window, a realistic cost assumption, or a check on which trades produced the return. Cataloged by failure mode rather than by indicator.
I run an algo bot 24/7 but I’m not always at my desk, so I built a pocket-sized thin client to watch the market and step into trades from anywhere — an 8-inch mini-laptop running Lubuntu and Tailscale.