Skip to content
← All work

Vericast · USA · Backend AI Engineer · 2025 – Present

Vericast Marketing Strategist (vAI)

A production agent that lets financial marketers hold a strategy conversation with an autonomous AI.

AI / LLM

A production agent that lets financial marketers hold a strategy conversation with an autonomous AI.

~$946 / 2wk
LLM token waste eliminated
PythonFastAPILangGraphLangChainClaude (AWS Bedrock)React 18PostgreSQLRedis

Overview

A production conversational-AI platform for financial-institution marketers. The agent runs SQL against a data lakehouse, browses the web, executes sandboxed Python, analyzes files, and renders interactive React data-viz artifacts inline in chat — built on a LangGraph THINK → ACT → OBSERVE loop over Claude via AWS Bedrock.

Approach

01 — Architecture

Main agent calling 30+ LangChain tools plus dedicated subagents (SQL query, artifact builder, web analyzer) in a multi-service monorepo: React frontend, FastAPI backend, sandboxed code-executor microservice, PostgreSQL, Redis.

02 — Reliability

Real-time agent-loop guardrails framework — a HookedToolNode wrapper + a HookRunner orchestrating pre/post-tool hooks with a 4-action PASS/RESOLVE/REWRITE/OBSERVE model — that detects pathological tool-call loops and self-corrects.

03 — Observability

Self-hosted Langfuse via a custom REST callback handler: per-request / per-session / aggregate cost tracking against a Bedrock pricing table, latency monitoring, and PII redaction (SSN / account-number scrubbing) keeping 100% of sensitive data on-prem.

04 — Security

Hardened the sandboxed Python code-executor, closing two vulnerability classes (sandbox-escape, path-traversal) with a Bruno security test suite.

Results

~$946 / 2wk
LLM token waste eliminated
1M+ tokens
Context blowouts prevented
1,500+
Backend tests authored

What I owned

  • ↳Designed and built the real-time guardrails framework (HookedToolNode + HookRunner, ~700 LOC, 86 tests, zero regressions) that eliminated a recurring ~$946/two-week LLM token-waste problem and 1M+ token context blowouts.
  • ↳Built self-hosted Langfuse LLM observability via a custom REST callback handler (~500 LOC) with per-request cost tracking, latency monitoring, and PII redaction — keeping 100% of financial data on-prem.
  • ↳Refactored the web analyzer into 1 of 4 internal LangGraph subagents, removing a cross-service network hop and adding a 'Quick Analyze' fast-path that fixed timeouts on heavy SPAs.
  • ↳Authored the backend test suite (1,500+ tests) across agent orchestration, tools, database, and auth — the platform's regression safety net.

Next project

Imperium — Trust Management Platform

→