HomeCase studies

Results, measured.

Real systems, real numbers, measured the boring way: only what was verified counts.

Featured · Lead generation

Real estate lead generation

A real estate team needed a reliable pipeline. We mined public permit filings and search databases to find owners who were likely in the market, then verified and qualified every record with AI before it counted.

approximately $1.09per verified lead

Verified leads
230
Awaiting verification
+273

Leads still pending verification are not included in the cost per lead.

How the system works

Most records never make it past verification. That's the point.

  1. Step 01

    Source

    Public permit records and search databases, mined for in-market owners.

  2. Step 02

    Verify

    AI cross-checks and qualifies every record before it counts.

  3. Step 03

    Store

    Verified leads land in a clean database, ready for the team.

  4. Step 04

    Report

    Cost per verified lead, tracked live.

More work

Other things we've built

Products, client systems and the occasional rescue job. Clients are described rather than named. Tap any one for the detail.

Product · Real estate developersIn pilot

10laws, rules and regulations, searchable clause by clause

Every clause, checked against the law.

Khushi: contract drafting grounded in RERA. Drafts in the developer's own wording, on their own servers.

  1. Upload a draft
  2. Pull the law
  3. Draft with sources
  4. Check every citation
  5. Lawyer signs off
Read the full write-up

Developers draft agreements for sale, leave and licence agreements and redevelopment documents over and over. Generic AI tools draft in US legal language, invent citations and send the text to someone else's cloud. Legal teams need drafts in their own house language, with every reference traceable to an Act or rule.

  • A drafting and review workspace: chat with a document, review it clause by clause, apply suggested edits, and run reusable reviews across a set of documents.
  • A regulatory layer of 10 instruments, searchable clause by clause: the RERA Act 2016, the Maharashtra RERA Rules 2017 and their 2019 amendment, five MahaRERA regulations, the Registration Act 1908 and the Maharashtra Stamp Act 1958. Older versions are kept and marked as superseded.
  • A client layer: each developer's executed contracts load into their own instance, so drafts follow their wording and their playbook.
  • Citation checking: every quoted passage is matched against the source document before it is shown.
  • A Microsoft Word add-in (beta), so lawyers can work from the document they already have open.

Each client gets an isolated instance on their own infrastructure. Which model it uses is a setting the client controls, including fully local models. A fully local set-up sends nothing outside the client's network.

Built and in pilot. Security review findings are closed in code and waiting on a live test. We are looking for one developer to run it on real agreements.

  • Next.js
  • Express
  • Supabase
  • S3 compatible storage
  • Docker

Lead generation · Real estateBuilt and tested

20cities watched for new projects, 11 liveWho is about to build, and who decides.Signal Radar

The first version ran as a large visual workflow. It produced verified leads, but it was hard to change, ran out of memory on big batches and could not be handed to a sales team. And projects are often filed under a single-purpose company, so the name on the permit is rarely the company that makes decisions.

  • Collectors for public building and property data in 20 cities, 11 of them live today, plus news-based discovery for the rest.
  • A ledger of every project already seen, so nothing is processed twice.
  • A developer resolver that looks past the permit's company name to the firm behind it, and keeps the evidence for each answer.

Built and tested, with over 1,000 automated tests. Live runs in New York, Toronto, Austin and Miami, each with a spending cap. Next: daily scheduled runs and more cities.

  • Python
  • Public open data
  • Search and language model APIs
  • Local web app
  • Excel export

A real estate fund · Investor relationsBuilt, first deployment in scoping

0emails sent without a person reading themRemembers every investor. Drafts the follow-up.Investor relationship cockpit

Senior partners at funds and family offices meet dozens of investors a month. Follow-ups slip because the context sits in someone's head, on a business card or in an old email. A spreadsheet helps, but it doesn't remind anyone or write the email.

  • A record of people, organisations, meetings and notes, entered from business cards and after each meeting.
  • A morning job that lists who is due a follow-up: enough time since the last meeting, a priority contact, or a meeting coming up on the calendar.
  • AI drafting in one step only. The model sees one contact's details at a time and writes a short draft in the partner's tone. It cannot send, delete or browse.

Built and tested, with 36 automated tests covering data isolation between firms and draft safety. The first deployment is being scoped for a real estate fund, inside the fund's own cloud and with a model the fund already approves.

  • Python
  • A database per firm
  • Scheduled jobs
  • An approved language model
  • Outlook drafts
  • Docker

A real estate fund · ReportingIn use

21voice notes in, three reports out, same dayTalk on site. The report writes itself.

Funds that lend to developers send someone to check each site: construction progress, labour on site, sales office activity, access roads. The visit takes an hour. Writing it up in the fund's format takes much longer, and notes taken on a phone at a noisy site are patchy. Three sites in a day meant three reports to type.

  • A report structure taken from the fund's past reports: header, attendees, location, a site update, layout plan and a captioned photo grid.
  • Two checklists, one for land-stage sites and one for sites under construction, so each visit covers the right points.
  • A drafting step that turns a day of recorded voice notes into one report per site, in Word.

Used on a real day of visits: three sites, 21 recordings, three reports. Packaged so the next visit follows the same process.

  • Claude
  • Wearable voice recorder with transcription
  • Word document generation

A storytelling media account · Content automationLive

60short videos a month, nobody editingScripted, voiced, cut and posted on schedule.

Short-form accounts grow by posting every day. By hand, that means writing a script, recording a voice, cutting background footage, adding captions and uploading, for every video. Off-the-shelf video tools charge per render, which adds up at daily volume.

  • A script writer that picks a few random keywords and writes a short story with a strong opening line, plus a short caption.
  • A voice step that turns each script into narration, with word-level timings.
  • Our own rendering service in place of a paid video API: narration over looping background footage, word-by-word captions, a finished vertical video.

Live and posting. The renderer uses set edit templates rather than AI video generation, which keeps each video cheap to make.

  • n8n
  • Remotion
  • Express on Google Cloud Run
  • Whisper
  • Text-to-speech
  • Claude and Grok
  • Google Sheets and Drive
  • Meta Graph API
  • YouTube API

Infrastructure · n8nDone

11 → 0workflows at risk, across 237Upgrade-ready in an afternoon, nothing stopped.n8n 3.0 readiness

n8n 3.0 removes old nodes and changes several server defaults. Workflows that use a removed node stop running, and published ones stop activating. Teams that self-host need to know what will break before the upgrade, not after it.

  • Ran n8n's migration report: 11 workflows affected by 4 workflow rules, plus 4 server-level changes.
  • Used the automatic migration where it was safe, then checked each changed workflow for nodes whose output had changed shape.
  • Rebuilt the rest by hand: two old Function nodes moved to Code nodes, and four AI agent tools rebuilt on the current HTTP Request node, keeping their credentials and descriptions.

Reviewing every workflow surfaced credentials typed straight into node settings, and a disconnected step that meant one agent's results never reached its sheet. Both were flagged to fix.

Done. The instance is ready for the stable 3.x release.

  • n8n queue mode
  • Docker Compose
  • Postgres
  • Redis
  • n8n MCP server
  • Claude

Want numbers like these?

Book a free 30-minute call and we'll look at where your pipeline leaks, and what it would cost to fix.