Eric Lazarus
Systems that have to be right and have to be fast — and the numbers that prove it
Everyone in finance is currently either bragging about AI or panicking about it. I’d rather just show you what happens when someone who’s spent years inside kdb+ trading infrastructure actually points AI at the problem, instead of bolting a chatbot onto a pitch deck. Two things below. Both have working code and a benchmark attached.
Two things I build
LBT → A load-balanced kdb+/q query dispatcher that gets you to 24-hour operation without re-architecting what you already run — canary deploys, live routing edits, degradation alerts before your desk feels them, and a tickdemo that hot-swaps a day’s data with zero downtime. Benchmarked at 10x-or-better elapsed-time improvement from locality-aware routing.
CLEAR → Five distinct AI investment-philosophy evaluators — Munger, Buffett, Klarman, Lynch, Simons — each with real kill-switch logic and its own scoring model, not one model roleplaying five personalities. Full production stack, an MCP server exposing the analysis as tools, and a semantic consistency pipeline that’s 10–20x faster than checking it by hand.
Why believe any of this
Because the pattern isn’t new for me — it’s the same move I’ve made for over a decade: go into a system that’s slow, opaque, or fragile, and leave it measurably faster with the numbers to show it.
75% backtest time cut — 3 hours to 45 minutes, via predictive, data-local load balancing
0 data-leakage findings in an internal audit of a live trading-strategy pipeline
22 min mean time to recovery on the platform I currently lead
Right now I’m the Lead Architect for a 3-person kdb+/q team building a system that turns market narrative — newsletters, filings, text — into validated, beta-neutral trading strategies: 85 candidate patterns explored per hour at a P95 shard latency of 70 seconds, idea-to-decision time cut from five days to five hours, 26+ text-ingestion adapters feeding it. Before that: specifying two risk-management and trading-desk systems at Millennium Management; building the sole analytics dashboard behind a 117,000-person training program at Verizon Wireless; and, earlier in my career, architecting a Smalltalk-based derivatives trading system at JPMorgan Chase that MIT Sloan Management Review later cited — in “The Case for Expressive Systems” — as one of the most financially successful systems ever built in the language, eventually adopted by thousands across the firm, and later profiled again in a Cincom case study.
Around all of that: a philosophy degree from Columbia, six years of hands-on kdb+/q, fifteen-plus years of full-stack engineering before that, and co-founding CodeBorging, a 500-plus member community of engineers using AI and LLMs to get better at programming — several of them trading and vector-programming specialists. Also shipped: bddq, an open-source AI-assisted BDD testing tool for q.
I’m not interested in being the loudest voice in AI-and-finance. I’m interested in being the one whose claims survive someone actually checking them.
Running kdb+, thinking about AI in finance, or just want to argue about whether any of this holds up? Reply — tell me what you’re working on.