Portfolio Agent

The problem

A portfolio can claim seniority; it is harder to prove it. The Portfolio Agent turns the claim into a working system: a grounded, cited chat that only answers from a real corpus, refuses everything off topic, pulls live commit data, traces every reply in public, and publishes its own accuracy score so nothing has to be taken on faith.

Stack

Screenshots

The full PRD lives in the repo. Open it above.

The Portfolio Agent runs as a LangGraph.js graph with four named nodes. The guardrail declines anything off topic or any prompt-injection attempt in a single redirect sentence, retrieve does a top-6 pgvector cosine search over the embedded corpus, the agent node runs deterministic tools (real GitHub commits, the published eval score), and answer composes a grounded, cited reply. Cost is engineered, not hoped for: all generation flows through OpenRouter on a prepaid key hard-capped at $5 a month at the billing layer, the visitor picks the answer model from a server-enforced allowlist, and if the credit is ever exhausted (an HTTP 402) the graph degrades to the Groq free tier with an honest caption instead of an outage. Every reply carries a public Langfuse trace with one span per node. The accuracy is measured, not asserted: a 40-question golden set (facts, live tools, guardrail probes, and Spanish) is graded by a free cross-family judge and published as a real table. The honest part is the iteration. The score moved from 92.5% to 100%, and every early miss was a judge limitation caught and fixed through a human review of disagreements, never a fabrication by the agent, whose anti-fabrication behavior is architectural: it answers from real tool output or says the lookup failed. Built in public with the agentic-dev-kit, one backlog item at a time, spec first.

Evaluation

The agent is measured, not asserted. A golden set runs on every build and the results are published here, straight from the committed JSON the agent itself reads.

Timeline

Live from GitHub, as of Jul 7, 2026