As a bootstrapped startup founded by our core team (and nope, we definitely don't come from wealthy families!), we don't have the luxury of deep-pocketed funding or instantly hiring amazing talent to scale fast. It’s our own wits, time and money at stake, which makes smart decisions around technology, frameworks, and services absolutely crucial.
When you're bootstrapping software or digital products, choosing the right tech is essential. You need to optimize your time, budget, and speed to market while still planning for scalability- which doesn't always neatly align with moving fast. Making poor decisions can mean spending precious time on the things that don't move the business in the right direction and incorrect use of cashflow that could otherwise fuel growth.
I'm going to share some key architectural and technology choices we made early on at Chirp AI, allowing us to quickly get up and running, keep operational costs low, and a tight focus on creating customer value and generating sales.
Here are some guiding principles that have consistently shaped our approach when building our architecture and choosing technology for building quickly, inexpensively, yet effectively positioning us to scale:
- Go serverless. Let someone else handle infrastructure management; pay only for what you use.
- Maximize cloud provider free tiers. Optimize your usage and you'll be surprised how far these tiers can stretch.
- Avoid reinventing the wheel. Just do a wee bit of extended research to find existing packages or solutions that meet your needs.
- Use popular, community-supported frameworks and technologies. Avoid the new shiny car as it avoids potential future hiring challenges.
- If you can, use click-ops to quickly deploy production-ready solutions, No shame as an engineer - just codify it later.
- Integrate AI into your workflows. Great power, great responsibility - use it wisely.
- Pragmatic engineering: Focus on what really matters, be realistic, sensible and solution-oriented - solve for customer value first.
- Right size your solution design: Enterprises require enterprise solution designs and start ups require start up solution designs, don’t over bake it.
Virtual Cloud Infrastructure
We chose AWS. While AWS might not be always the cheapest option, we firmly believe the benefits outweigh cost concerns. AWS offers an extensive, integrated ecosystem of cloud services, making it straightforward to build secure applications without significant time investment. Plus, it's backed by a vast community and plenty of serverless, pay-as-you-go options.
All our AI agents are run on servers managed through Fargate ECS instances on modest hardware specs. I won’t disclose the exact cost, but if you’re across general pricing then you know it's minimal, and the best part— essentially zero server maintenance overhead. This lets our team focus on engineering tasks directly tied to generating customer value.
Initially, we didn't waste time setting up infrastructure as code (IaC) or elaborate CICD pipelines. Instead, we went straight to the AWS console and set up infrastructure as needed. While some infrastructure resource changes require AWS CLI calls we simply resolved that by performing CLI only changes through the IDE terminal. We've now codified certain parts, mostly refining our information architecture and naming conventions to maintain resource and information organization as we continue to scale.
Data & Analytics Storage and Processing
Our main data source is customer calls and agent interactions. Using AWS S3 for raw data storage was an obvious choice—it's the cheapest form of object storage, requires zero maintenance, and integrates seamlessly with AWS Lambdas for event-driven processing. Each S3 object is small, so low-memory allocated Lambdas handle data processing efficiently. Thanks to AWS's generous Lambda free tier, our raw data storage and processing incur almost no cost!
There are no operational overheads for maintaining our event driven data pipeline, and we essentially only pay for actual usage although, technically, it’s mostly free. We’re also integrating LLMs within our data processing Lambdas via API calls, enabling complex unstructured data and text analysis, transformations, and deriving new data effortlessly. If you're clued up in the ML and data engineering space, performing data operations with intent to extract out coherent and high value insights through unstructured raw data has traditionally been a tricky process to do well and efficiently.
Processed and aggregated data gets written directly into our database (coming up next) for final cleaned data analytics storage. Our entire data processing pipeline, from call completion to consumable data, runs in near real-time with minimal cost and maintenance.
Customer Web Application (AI Agent Analytics Portal)
When building the frontend, we wanted to spend maximum time on customer-facing features, while ensuring future scalability and maintainability. We chose Next.js, a React-based framework, because it provided everything needed for a robust web application without custom-rolling advanced web features. Plus, its community support is huge, with 5 million+ weekly NPM downloads – it was an obvious choice.
For deployment, we went with Vercel, creators of Next.js as our native web hosting and deployment platform. Vercel manages hosting, deployment, CDN, and backend logic without a shmick of server maintenance and it's integrated seamlessly with GitHub CICD. Given this is an internal customer portal with low traffic and very low pricing, it's extremely cost-effective and scalable.
Lastly, you need a backend to your frontend. We paired this with Supabase, a backend-as-a-service powered by a fully managed PostgreSQL database — textbook players are going to throw stick at me me here but we took a pragmatic unorthodox approach with using an OLTP (online transaction processing) web backend database to play a dual role acting as our structured data analytics database. This works for us because this database is serving internal web portal requests which are currently very low traffic meaning that performing our occasional analytical queries won’t impact our customer experience — and we don’t pay for the extra compute or storage if we were to use an OLAP (online analytical processing) database avoiding the added cost and potential complexity of a separate OLAP database.
Alerting and Monitoring
All good engineering teams must quickly respond to incidents to minimize customer impact. Thus, all mentioned technology and service choices offer built-in monitoring and observability.
Vercel and Supabase provide out of the box analytics and observability, allowing alerts for unexpected traffic spikes or anomalies. Our AWS services all integrate seamlessly with CloudWatch, providing detailed minute-by-minute custom monitoring.
We also leverage 'Pino logs' node package within our agent application for production log levelling framework and structured JSON outputs to CloudWatch, simplifying parsing log data, supporting low cost and effort log analytics and monitoring. We’ve also setup Slack webhook integrations for immediately notifying our team of any triggered alerts from AWS, Vercel or Supabase that are verbose in detail, helping us stay customer-focused without unnecessary operational distractions unless alerted to.
Monitoring costs are minimal or included in our managed services pricing, and AWS CloudWatch incurs only modest fees as a large portion of CloudWatch data processing costs are absorbed within the free tier consumption.
Integrating AI into Workflows
As an AI startup, using AI is a no-brainer for us. We actively encourage our team to leverage AI tools, it greatly enhances productivity and arguably feels like a 3x force multiplier.
We leverage AI to generate detailed PR summaries, empower engineers with GenAI tools integrated directly into their IDEs to accelerate development, and automatically produce unit tests - streamlining a typically tedious process because no one likes doing unit testing! AI also drives the rapid creation of baseline product and technical documentation, API specifications, and a range of other essential documents, enabling the team to focus on real customer problem solving and value generating tasks.
However, we must use AI responsibly. AI-generated code is always reviewed thoroughly and adjusted to reflect specific broader business considerations and context that would be typically hard for an AI assistant to consider. Our team is also encouraged to contemplate on how they could further harden the code and error management to make AI-generated code more robust and reliable, something that AI code assistants don’t usually perform well in. AI shouldn’t replace well thought out code, rather it’s a valuable assistant that enhances our productivity — vibe coding and documentation is bad, AI pair programming and guided documentation is good.
Closing recap
Bootstrapping a startup means limited resources and the need to initially spend your own time and money optimally if you want to increase your chances of going to market at speed and creating business momentum to carry you forward. In this blog, I shared key decisions we made early on to minimize operational overhead, maximize speed to market, and scale affordably. Serverless infrastructure, cloud free tiers, pragmatic engineering choices, and integrating AI in appropriate parts of our teams operational and development cycle are very key to helping keep us focused on creating customer value. I dived into our key service choices for AWS, Next.js, Vercel, Supabase, and how we handle monitoring and alerting to minimise our teams’ active infrastructure operational overhead. If you’ve made it this far - thanks for reading!
