arrow_backBack to What I Build

Agentic Systems & Memory

Your company forgets everything it learns. So does your AI.

Off-the-shelf AI treats every prompt as a first date. I build systems that remember: longitudinal memory across projects, research agents that argue until the answer is solid, and orchestrators that route work to the right tool. The capability is memory; the business result is a team that stops re-explaining the same context every week.

What Usually Looks Normal Until I Point At It

Every ChatGPT thread starts from zero knowledge of your company

Research assistants accept weak answers because they cannot argue with themselves

Agent work is opaque — you cannot audit why a decision was made

Multi-agent demos look smart but never reach production

How I Build It

01

Start with memory: index what your organization actually knows and asks

02

Add adversarial critique so agents attack weak claims before you see them

03

Route tasks to the right model family for the job

04

Deploy local-first with audit trails, not black-box chains

Outcomes

Longitudinal semantic memory across sessions

Cross-family adversarial research that beats single-model critique

Technology Stack

PythonSQLiteReasoning-grade embeddingsLocal LLMsMulti-agent orchestratorSlack APIs

Capabilities

Longitudinal Semantic Memory

A local-first semantic memory system indexing longitudinal sessions. Query by project, date, source, or mechanism — not just keyword match.

How It Works

  • Process conversations and documents so the AI actually understands the meaning
  • Save everything locally and search it in a way that surfaces the most useful details first

Outcomes

  • Sub-second recall across 18+ months of work
  • Cross-project pattern discovery

Adversarial Deep Research

Most AI stops checking its work. I make separate research passes attack the evidence, contradictions, and missing synthesis before you ever see the report.

How It Works

  • Writer model drafts the report with predicted source claims
  • Cross-family critic attacks evidence, contradictions, and missing synthesis

Outcomes

  • 100+ inline citations on a head-to-head where the incumbent tool produced 43. 10+ primary sources — SEC filings, Senate records, government documents — where the incumbent surfaced zero.

The research is public.

Benchmarks run on machines I own, measured against frontier systems — with the failed approaches reported next to the results. Read it before you take my word for anything.