AlwaysCool.ai moved from spreadsheets and custom GPTs to LangGraph-based microservices after synchronous workflows began taking five to fifteen minutes.
2
OpenTelemetry gives the team shared traces, metrics, logs, and business-process data for compliance workflows and agent management.
3
LangGraph state management lets long-running workflows checkpoint progress, while human review remains part of the validation process.
Summary
Colin McNamera describes how AlwaysCool.ai introduced AI into food product formulation, nutrition analysis, label generation, and FDA compliance work. The team began with spreadsheets, ChatGPT, custom GPTs, and Python adapters to FDA and USDA interfaces. As models became more capable, synchronous requests started taking five to fifteen minutes, which broke their applications. They moved to LangGraph for state management and asynchronous flows, then began porting the workflows into Python services behind FastAPI. OpenTelemetry now connects their web applications and agent workflows to collectors, dashboards, and observability systems. McNamera argues that the resulting traces matter because the company needs to audit what an agent did, when it did it, and which business data informed the result. He is open about the system still being under development, and he keeps people involved through manual checks and review by a CPG lawyer.
Food compliance work started with spreadsheets and manual checking
AlwaysCool Brands develops reformulated products for retail grocery stores. During early bench testing, formulas and nutrition facts changed often, while labels still had to capture allergens, religious restrictions, and FDA requirements. McNamera says preparing the information correctly could take hours, with serious penalties for mistakes. The team initially handled this work through spreadsheets and manual processes, including checks by McNamera and a nutritional scientist.
Custom GPTs helped automate label generation, but open-ended consoles were hard to control
The team began by copying nutrition data into ChatGPT, then built custom GPTs for internal use. A Python adapter connected those workflows to FDA and USDA interfaces so the system could generate nutrition facts labels more consistently. They also put the tools into a private internal console using Open WebUI. McNamera says the approach helped, but console workflows made it difficult to guide delegated work and get the same process performed everywhere.
Long synchronous model calls forced a move to asynchronous workflows
The team built scripts and an AI agent that pulled labels and checked FDA regulations. After moving to more capable models, their applications changed from returning results synchronously to making users wait five or fifteen minutes. That broke the existing applications and prevented the team from moving the work beyond the lab. They introduced LangGraph TypeScript for state management in two applications that needed asynchronous processing.
LangGraph checkpoints let workflows resume after intermediate steps
McNamera describes LangGraph as a way to share state across nodes and agents and checkpoint a workflow as it runs. If a process stops after a file operation or an EDI-system request, the team can identify the step where it stopped and inspect the history. The company is porting its LangGraph TypeScript services into Python graphs behind FastAPI so it can centralize governance and keep the agent logic separate from the service layer.
OpenTelemetry turns agent activity into auditable business data
The team uses OpenTelemetry to instrument agentic flows and record how compliance work is performed. McNamera wants the logs to include business-process information and the KPIs observed by the agent. Their shared instrumentation reaches both web applications and agent processing, with collectors that can transform data, clear personally identifiable information, and apply governance rules. The resulting dashboards support compliance work and the operational management of agents.
Specialized services still require human validation
The emerging architecture gives each specialized agent an entry point through a small FastAPI service that boots a LangGraph. McNamera says the team validates specializations through direct observation, evaluation, and review with a consumer packaged goods lawyer. The dashboards provide feedback about the agents, but observability is not presented as a replacement for people. He says it is important to keep a human in the loop.
The team connects AI observability with conventional metrics, logs, and traces
In the discussion with Demetrios Brinkmann, McNamera describes routing agent data through OpenTelemetry collectors into Prometheus, Grafana, Loki, and Tempo. For AI-specific inspection, he mentions Langfuse, LangSmith, and Arize Phoenix. This setup can carry prompt versions, evaluations, flow inspection, and ordinary service telemetry into the systems the team already uses. He also mentions ClickHouse as a metrics database.
The first workflows became shorter and more repeatable
McNamera says LangGraph helped reduce jobs that previously took two to three hours to five to ten minutes. He connects that improvement to the move from console experiments to microservices, explicit state management, and observability. The team is still learning, since the software is pre-1.0 and may introduce breaking changes, but two applications have been deployed in the new setup.
"The challenge is that when you're creating a sample to send off to your customer overnight, some of your customers might have nut allergies or they may have religious preferences."Colin McNamera01:55
Who should watch
You are moving an LLM workflow from a private console or script into services that need asynchronous execution and shared state.
Your team has to explain what an agent did in a regulated process, including the inputs, steps, timing, and committed changes.
You are deciding how to connect LangGraph or another agent framework with ordinary logs, metrics, traces, and AI-specific evaluation tools.
