Agent systems I have built

Built an internal workspace where teams can create, run, and evaluate reusable agent skills. It supports live long-running work, company context from Jira and Confluence, and a three-layer evaluation system that measures execution quality rather than relying on model opinion alone.

Request analytics service streaming daily cost trends, per-model breakdowns, and usage comparison across all AI API calls. Token-counting service built from scratch: tiktoken with the OpenAI Vision token formula (image scaling, tile computation, detail modes), pre-request estimates validated against actual API usage.

It sells a Company Enrichment API, fulfilled by a multi-agent research pipeline: parallel domain researchers under a central orchestrator, entity disambiguation before any expensive call, iterative knowledge-gap loops. The business runs itself around it: customers pay retail through an HTTP 402 paywall, and revenue and costs settle through an Agent Wallet.

The hard part wasn't proving an agent can run a business, it was making one you'd trust to leave alone: a spend gate on every payment (daily cap, runway floor, fail-closed approval above a threshold), decline-before-charge on unprofitable work, and a per-job P&L ledger recording every decision on-chain.

Every tool call is checked against a policy engine before it runs. When a call crosses a line, the gateway holds the connection open and waits for a human. A context server exposes the company's own rules and precedents to every connected agent, so decisions are made against institutional knowledge rather than a generic model's guess. Every decision is written to an audit trail that captures the full context, inputs, matched rule, grants evaluated, and who resolved it.

Runs on my own server, from finding opportunities to drafting proposals, building tailored CVs, and filling forms. Drafts use documented experience and pass automated checks and a separate model review.

Sandboxed drafting, tracked workflow state in Postgres, and a fixed CV template keep each step controlled. Human approval gates irreversible actions; submissions are verified against the platform's confirmation.

Weaviate hybrid search (BM25 + dense) with configurable alpha, query expansion via synonym generation, post-retrieval reranking with time-decay × priority-boost. T5 fine-tuned on synthetic task scenarios, 4-bit quantized, deployed on Cloud Run with GPU.