Skip to content

// THE WORK

Real work. Real numbers.

We don't show you polished case studies. We show you what we actually built, what broke, and what the numbers look like.

LegalDelivered

Running daily, spending money, writing zero rows.

A legal intelligence platform was running daily, spending money, and writing zero rows to its database. Wrong data source. Wrong endpoints. Wrong cost model. We rebuilt all three.

5

Root-cause bugs found

3

Counties covered after rebuild

0

Rows before the rebuild

Read the full case study →

DistributionExpanded

101,862 properties. We found our own waste and reported it.

Broken dedup queries and backwards name parsing. We documented every mistake, and the client expanded the engagement.

101,862

Properties in the database

~40%

Share enriched

Expanded

Engagement result

Read the full case study →

EnergyProduction

77% of ZIP codes over a federal legal limit.

A water quality product that nobody wanted, until we stopped leading with the technology. Across roughly 300 Connecticut ZIP codes, 77% had a contaminant above its federal legal limit.

~300

Connecticut ZIP codes analyzed

77%

ZIPs over a federal legal limit

233 of 301

Recount, independently

Read the full case study →

Enterprise SecurityOngoing

Country-level risk. Enterprise exposure. Decision-ready in under 60 seconds.

HPE needed to know whether their people and facilities were safe, anywhere in the world, at any time. We built the intelligence layer that makes that possible.

Worldwide

Coverage

Dozens

Data sources integrated

<60s

Time to decision

Read the full case study →

InsuranceOngoing

Five platforms. One operational picture.

Data scattered across five systems. Zero unified customer view. We built the operational picture, and then stayed to build everything else.

5

Data sources unified

None

Unified customer view before

Ongoing

Status

Read the full case study →

Safety TrainingDelivered

1 million people trained. Zero digital presence to show for it.

A workplace behavioral safety firm had trained over a million people across many years and many sectors. Their website told none of it. We changed that.

1M+

People reached through content

Multiple

Category rivals cited by AI

1,175 lines

Schema markup written

Read the full case study →

// WHAT WE SHIPPED

Things we built. Go use them.

Case studies are easy to write. These are live products you can open right now, including the one where we caught ourselves publishing data that wasn't real.

ProjectKids

Consumer marketplace, built and operated by us

An activity and camp platform that scans over a thousand program sites daily and turns them into a searchable, bookable directory across three metros.

It is also where we graded our own data: a confidence spine that scores every listing by how it was obtained, scraped, inherited, or model-derived, and serves it by trust tier instead of quietly hiding the weaker rows. That work established how many listings were genuinely bookable, meaning a real future session and a real price: 906 of 1,290.

projectkids.io

Edwin

Our AI memory and verification research

An AI memory system that verifies what it stores and challenges what it believes, using multiple models arguing against each other to surface the truth.

43,000 lines of Python, with 2,827 of 2,842 tests passing. A provisional patent application (USPTO 63/991,739) with 10 claims validated against 35 experiments. Two of those experiments disproved what we expected, and the filing says so.

heyedwin.io