AI is ready.
Are your operations?

The hard problem: can an LLM-driven system resolve its own tendency to hallucinate, design and backtest its own trading strategies, learn from its results, and deliver consistent performance against something as unpredictable as the stock market? Designed and built a fully autonomous multi-agent paper trading system that addresses it end-to-end — scanning the market nightly, generating and backtesting its own directional strategies, executing trades with no human in the loop, and refining its approach based on outcomes. Discrete agents own data ingestion, signal generation, position sizing, and risk management. Live trade activity streams twice daily to a public Discord channel for external verification.

Frontier LLMs are expensive, slow, and overkill for narrow tasks. Small, well-tuned specialist models can match or exceed them on their home turf — at a fraction of the inference cost, latency, and token burn. The hard part is keeping a compact model sharp over time without human intervention. Engineered a closed-loop autonomous fine-tuning pipeline that solves it: the model's own inference results drive each subsequent tuning cycle, with no human in the loop. The outer-loop system tracks hypotheses, constructs datasets from live outcomes, runs parameter-efficient training via LoRA adapters, debriefs results, modifies its own tuning strategy based on what worked, and redeploys the new checkpoint. Rinse and repeat — the trainer learns from the results of its own work and adjusts the next cycle accordingly. Recursive self-improvement applied to directional signal generation.

Taxes and LLMs posed an interesting challenge. While models may hallucinate, the IRS doesn't. Designed a workflow that lets an LLM take in volumes of IRS publications and read them (something LLMs are good at) and then self-direct its way to a rules-based engine that runs multiple loops of trial and error until the attributes and conditions it designed are 100% deterministic (no inference) and consequently 100% accurate. The result: a rules-based codebase that requires zero tokens to operate and can not only do your taxes, but find deductions and favorable tax treatments you never knew existed.

Citizens who have served their time face an exceptionally complex reality — a maze of eligibility rules, records, jurisdictional quirks, and support organizations that vary by state, county, and conviction history. A single missed benefit can mean returning to the system. Built a rules-based engine that self-updates as laws change, self-learns from new cases, and adapts to an effectively infinite set of individual circumstances to guide each person to the resources they're entitled to and keep them out of the system. Compassionate design; protects a vulnerable and sometimes forgotten part of our society.

Coding agents are powerful when they can see the relevant code — and dangerous when they can't. Massive codebases break this assumption: any single file is one node in a web of cross-cutting dependencies the agent never has in context. The result is plausible-looking code that calls functions that don't exist, ignores APIs that do, and breaks invariants three files away. Designed an MCP server that continuously maps the full codebase — symbols, call graphs, import chains, transitive type relationships — into a queryable index of code-understanding artifacts with evidence pointers back to source. Any agent can pull just the slice it needs, with Observed / Inferred / Unknown confidence labels per claim, and a validator that rejects anything not traceable to a pointer. The context problem doesn't go away on large codebases — but the agent stops paying for it.

Open-source orchestrator for multi-workspace AI projects. Routes human intent to sovereign agent Leads, each operating inside its own workspace under a brief that defines what it owns, reads, and must not touch. Handles context-rich prompt construction, non-blocking background dispatch, structured result verification, and cross-team coordination. A working framework for running a team of AI agents as one coordinated system.

Open-source reference implementation of the patterns behind production multi-agent Claude Code systems: sovereignty briefs that encode safety and code-control boundaries, XML-tag prompt protocol, structured result contract, persistent session continuity, peer-message broker, DAG-based cross-agent dispatch, and health-remediation loops. A working blueprint for anyone building their own multi-agent system.