AI products for startups
A working AI product in weeks, on your own accounts, with a written report every week.
Blog / Category
AI systems
AI for sales, support and operations: assistants, scoring and automation that people can check and trust.
An AI demo takes an afternoon. An AI system that a sales team, a clinic or a fund manager relies on every day takes engineering: structured outputs, fallbacks when the model fails, human review where it counts, and a record of every decision.
We build AI agents and assistants that work with your own data and your own methods, and that people can check: every score with its evidence, every answer traceable to its source, every change tested before it ships.
This is for you if
- You want AI in sales, support or operations, not another chatbot demo
- Your method or data is your edge, and the AI has to use it, not replace it
- Your industry needs accuracy, security and an audit trail
Articles
Check a warning against a known answer before a manager sees it
Pipeline warnings that nobody has checked are just noise with confidence. So before launch we built a test account where we knew every right answer, and the product had to match all 23.
Value: the goal isn’t AI adoption
Nobody needs AI; they need recovered revenue, a coached team, a product that ships. How we pick the outcome before the model, and use AI only where it earns its place, with examples from a school platform and a traffic recovery.
Trust: a successful request is not a correct answer
An AI call that returns successfully can still be wrong. The checks we build around every model before people act on its output: structured outputs, rules alongside the model, fallbacks, provenance, review and tests.
Normalization: same customer, five IDs
Every system has its own idea of a customer, a note or a date. Normalization turns them into one validated shape, and it quietly decides what your AI will say. Lessons from an AI platform for lawyers and clinicians.
Ingestion: start where the business actually runs
The first stage of every project is getting the real data in, from where the business actually operates. Why a sample is not enough, and what a traffic diagnosis and a sales product taught us about it.
Trusted systems, not AI demos: the five stages behind every project
Ingestion, normalization, trust, value, production. The five stages every project of ours goes through, from a website losing revenue to an AI assistant for clinicians, and why the model comes fourth.
How one person produces thousands of lessons with AI
SuperEscuela has 733 topics, 2,786 lessons and 13,705 practice questions, written by an AI pipeline against Costa Rica’s national curriculum. How the pipeline works, and where the AI is kept out on purpose.
AI speeds the work. People stay responsible.
AI makes investigation, comparison and iteration affordable at a scale that used to be impractical. It also makes confident mistakes. How we split the work: what AI does, what stays human, and how we review what it produces.
A confident wrong number is worse than a missing one
A close rate that read 100% when the truth was 50%, feeding seven reports a sales manager would act on. How it happened, why we fixed it at the source, and what to check in your own dashboards.
The cost of an AI feature: tokens, caching and the right model per task
What an AI feature really costs is a number per unit of value: per lesson, per answer, per document. How to estimate it, get the quality right first, then bring the cost down with model routing, caching and batching.
Connecting AI to your tools safely with MCP
MCP lets one integration serve every AI assistant your team uses. How to expose your systems through it without handing a model more power than it needs: narrow tools, read-only first, approvals and a log of every call.
Prompts are code: version them, test them, review them
A prompt decides what your AI system does, so change it the way you change code: in version control, reviewed, tested against fixed examples, and recorded on every output so you can explain it and roll it back.
How to test a RAG system before your team trusts it
A RAG demo answers the questions its builders thought to ask. How to test retrieval and answers separately, check citations automatically and agree a pass mark in advance, so your team can rely on what it says.
RAG that answers from your data, not around it
Most RAG failures are data failures that retrieval exposes. What must be true before retrieval, from sources of truth and permissions to citations and hybrid search, so your assistant answers from your knowledge base.
Least privilege for AI agents and integrations
An AI agent can do exactly as much damage as its credentials allow. How to scope keys, tokens, tools and write access so your AI integrations stay useful and contained.
Let customers connect their own CRM in one click
Self-serve software cannot ask every customer to copy an access key and book a call with their admin. How we built a one-click, read-only CRM connection, and why “what you did not share” deserves a screen of its own.
Guardrails: decide what the AI must never do
Guardrails turn “the model should not do that” into rules the system enforces. How to set them on what goes in, what comes out and what the AI is allowed to do, and how to test and maintain them like code.
Governability: who changed the prompt, and can you roll it back?
An AI feature is governable when you can say which prompt, model and settings produced any output, who approved that version, and how to go back to the previous one. How to build that in from the first release.
Human in the loop, placed where it counts
Human review makes AI safe to act on, but only when it sits at the right points and people can actually do it well. Where to place reviewers, how to design the approval step, and how to turn every correction into evidence.
Ingestion, normalization, trust: feeding practitioner data to an AI assistant
An AI assistant for legal and healthcare practitioners is only as good as the records it can read. What we learned owning the integration and data layer of such a platform, and why calling an endpoint is the easy part.
Observability for AI features: what to log, trace and watch
When an AI feature misbehaves, you need to know which request, which prompt and which model did it, and what it cost. What to log, trace and watch so you can answer that in minutes, not days.
The right model, not the cheapest: what a few cents per lesson buys
SuperEscuela started writing lessons with the smallest, cheapest AI model. We measured the quality against what children deserve, and moved to a larger one. Why the saving was never worth it, in numbers.
Put one gateway between your product and the AI models
When every feature calls AI providers directly, keys, costs and personal data end up scattered across the codebase. One internal gateway gives you a single place to control them, and lets you change models without touching a feature.
The AI harness: the code around the model is what makes it reliable
A model call is easy to add and hard to depend on. The harness around it, from prompt templates and schema validation to fallbacks and provenance, is what turns a model into a feature people can act on.
Webhooks or polling: keeping your AI's context fresh
An AI assistant that reads yesterday’s data gives yesterday’s answers, with full confidence. How to choose between webhooks, polling or both, and how to keep the data your AI reads as fresh as the decisions it supports.
How to trust an AI score
An AI that rates leads, deals or documents is only useful if people can check it. The six things we build around every model before its scores reach a sales team.
Agent or workflow? When to let the model decide
Most AI features are more reliable, cheaper and easier to test as a fixed workflow with AI in a few steps. How to tell when an autonomous agent is worth its cost, with a checklist you can use on your next feature.
Scoring sales calls with AI, against your own scorecard
A sales-coaching client had a scorecard proven by hand. We taught a product to score real calls against it, with evidence quoted from every call. What made the scores worth trusting.
What the first eight weeks of an AI product look like
A sales-coaching expert had a method proven by hand, a clickable mockup and an 8-week plan. The core product was working in eight weeks, as planned. Week by week, what that took, and how risks were handled along the way.
Related work
An AI sales-coaching product: from clickable mockup to working product in 8 weeks
Sales technology · client details withheld
A sales-coaching expert had a method proven by hand and a clickable mockup. We built the product around six AI agents that score calls against the client’s own scorecard, watch the CRM pipeline, prepare managers, coach reps, run one-to-ones and diagnose dips, with self-serve sign-up, CRM connection and billing.
Sales technology
Case data for a legal AI assistant
Practitioner AI platform · client details withheld
Integrations with legal case-management systems for an AI assistant used by law practices: ingesting contacts, cases and notes through REST and OAuth2, and normalizing them into validated records the assistant could use as context, with retries, rate limits and traceable logs around every call.
Legal services
Patient data for a health AI assistant
Practitioner AI platform · client details withheld
Integration work with healthcare practice, health-record and wellness systems for an AI assistant used by clinicians: ingesting practices, patients, notes and test results across REST, SOAP/XML, API keys and webhooks, normalized into one validated model, and handled to HIPAA, SOC 2 and GDPR expectations.
Healthcare
A mutual fund management tool with an AI assistant
Financial services · client details withheld
A management tool for mutual funds, built together with an AI assistant for the team that runs them.
Financial services
Building an AI product?
You do not have to do it alone.
Tell us what you are building and where it is stuck. We will tell you what is worth building next, and what the first practical step could be.
Prefer to write directly? enable JavaScript to see the address