All case studies

AI app · Customer support

Support answers from your own docs, or an honest "I don't know".

Tickets are sorted by topic and urgency, answered from your own help articles with the sources linked, and sent by a support agent who can edit the draft.

For
Startups
Built as
21-Day Build
Next.jsPostgrespgvectorClaude APIIntercom
  1. 1Ticket arrivesFrom Intercom or the in-app help form.
  2. 2Sorted by topicThe model tags topic and urgency, and urgent ones jump the queue.
  3. 3Docs searchedOnly your help articles and past answers are searched for the facts.
  4. 4Reply draftedA draft that links the sources it used, or says it doesn't know.
  5. 5Agent sends itEdits are saved and become test cases for the next release.

  • 100%

    Answers sourced

    Every draft links the passages it came from.

  • Every

    Release tested

    Run against past tickets before it reaches customers.

  • 1

    Person check

    An agent sends every reply until it's accurate.

"Our prototype answers support questions in a demo, but it makes things up with real customers."

The problem

A support chatbot built in a weekend looks great in a demo. With real customers it answers questions it has no facts for, mixes up plans and features, and nobody can tell which change made it worse.

The support team stops trusting it, and the startup is back to answering everything by hand, with an AI bill on top.

What we build

Your help articles and past answered tickets are split up and indexed in Postgres with pgvector. When a ticket arrives, the model sorts it, the search finds the relevant passages, and the reply is drafted only from those passages, with links back to each one.

Every edit an agent makes to a draft is saved. Those edits become a test set, and every prompt or model change is run against it before it ships, so you can see whether answers got better or worse.

The checks built in

Answers come only from your sources. If nothing relevant is found, the draft says so instead of guessing.

Every change is tested against past tickets before release.

If the AI service is slow or down, tickets fall back to the normal queue and nothing is lost.

Typical stack

  • Next.js
  • Postgres
  • pgvector
  • Claude API
  • Intercom
How we build AI features for startups

Startups

AI features we buildfor startups.

Each one is built in production code, with the parts that keep it working once real users arrive.

  1. Document extraction feature

    ReactNode.jsS3GeminiZod
    1. 1File uploadedPDFs, scans or photos, stored privately against the user's account.
    2. 2Fields readThe model pulls out dates, names, totals and terms into a fixed format.
    3. 3Checked in codeEvery field is validated, so a wrong date or total can't get through.
    4. 4Unsure fields flaggedAnything the model isn't sure of is highlighted for the user to confirm.
    5. 5Saved to the appClean records the rest of the product can search and report on.

    AI app · Document processing

    EveryField validated

    Upload a contract, get clean data back.

    Users upload PDFs, scans or photos. The AI pulls out the dates, names, totals and terms into a fixed format, and anything it isn't sure of is highlighted for the user to confirm.

    How the document extraction feature works

  2. In-app data assistant

    Next.jsPostgresOpenAI APIRedisPostHog
    1. 1User asksA plain-English question inside the product.
    2. 2Permissions checkedOnly records that user is allowed to see are ever searched.
    3. 3Answer builtThe reply links to the records it came from, so users can check it.
    4. 4Costs cappedPer-user limits and caching keep the AI bill predictable.
    5. 5Usage trackedWhich answers helped and which didn't, to guide the next change.

    AI app · In-product assistant

    EverySearch permission-checked

    Answers from each user's own data, and nobody else's.

    Users ask questions in plain English inside the product. Answers come only from records they're allowed to see, with links back to each one and a cap on AI costs.

    How the in-app data assistant works

All startup case studies

Working in 21 days from access, or your build fee back. Half paid only when it works. Full terms

Book a free 30-minute call