Sparkworks

AI made a mess.
We put on a clinic.

Book a data therapy consult
Connecting the dots · enabling new forms of access

For teams whose AI got ahead of their data.

Our forward-deployed team embeds with yours, finds where things went wrong, and fixes them at the source. You come away with AI that can stand up to clinicians, scientists, customers, regulators, and your board.

CLEAN UP

Fix what you already have

We audit what you already run, from code and prompts to pipelines and vendor licenses, then restore the lineage, testing, and monitoring that make it trustworthy and enterprise-grade.

BUILD & DEPLOY

Ship new products safely

We design and build AI products inside the controls you have, on whatever cloud and models you already use. High governance and regulatory compliance go into the plan on day one.

TRANSFORM

Bring your people along

New tools change little if the organization around them stays the same. We train your people and help update policies and processes until the new systems are part of how the work gets done.

HOW AN ENGAGEMENT RUNS
01
Intake
What you have built, what is stuck, and what a 10/10 outcome looks like.
02
Diagnosis
A hands-on review of the tools and models, and of the people, processes, and policies around them.
03
Treatment
We fix, rebuild, or ship new work alongside your team, and we treat the holistic system, including the parts that are not code.
04
Follow-up
We keep watching after launch, maintain what we built, and improve it as your needs change.

Our healthcare data work, in more detail.

Three previous projects. Client names and exact figures stay private. Where we show numbers, they come from tests on synthetic data or are marked approximate.

HEALTHCARE DATA · CASE STUDY

Teaching a data agent the rules of the data.

A life-sciences team wanted its researchers to ask plain-English questions of real-world oncology data and get patient counts back. Most of the time the agent got them right. When it was wrong, nothing about the answer said so.

DOMAINReal-world data · HEOR
OUR ROLEDesign and build
DATALicensed oncology data products, inside the client’s own warehouse
STATUSAgent built; safeguards tested on synthetic data
2 → 16
Questions right out of 16 once the agent was given checked rules instead of bare column names
14
Wrong answers in the first round that came back with no error at all
yes ≠ True
One mismatch between a vendor manual and the data made whole queries return zero
600
Synthetic patients and 16 questions, with every right answer known in advance
TRY IT · BREAK THE AGENT, THEN FIX IT

Pick a question, then change what the agent is given. These are the real results from our test on synthetic data.

{{av.q}}
AGENT ANSWERED{{av.said}}
CORRECT ANSWER{{av.truth}}
VERDICT{{av.verdict}}

{{av.why}}

WHAT WE DID
01
Read the sources

Each licensed data product comes with a data dictionary and a usage manual. We read both and turned them into short, single-statement rules, each pointing back to its page.

02
Test every rule against the data

Before a person approves a rule, it is run against the real tables. Do the columns exist? Do the values match what is stored? Is it really one row per patient? This step is where the disagreement between the manual and the dictionary over values like yes and True turned up.

03
Give the agent only what is approved

For each question the agent gets the approved rules and the real column values for the tables it needs, and nothing else. A query that returns zero rows goes back as a failure, not as an answer.

Writing the SQL was never the problem. The agent did not know how this particular dataset records things: which values mean yes, which dates count, which rows to leave out. Once it did, 16 of 16 test questions came back right, with no retries.
AI EVALUATION · CASE STUDY

Can a small decision model check the agent’s work?

Jev, a new model from TypeSafe AI, does not write text. It picks from a set of options we give it, such as yes or no, and reports how confident it is. We tested whether it could review each plan our healthcare-data agent writes before any number reaches a user.

MODELJev, by TypeSafe AI
OUR ROLEEvaluation and design
TEST DATASynthetic only
STATUSExperiment. Nothing is in production.
14 / 14
Planted mistakes caught in oncology questions, with 1 false alarm in 16 correct plans
≈ 0.3 s
To check one plan, against about 3 seconds for a general-purpose AI
≈ 50¢
For ten thousand checks
2 → 11
Questions right out of 16 when Jev checked plans against the rules, up from 2 with a dictionary alone
PLAY · CATCH THE SLIP

These are six plans the agent wrote during our tests. Some are correct; others answer a slightly different question from the one asked. Read each plan, decide, then see what Jev decided.

{{jTag}}
{{jq.q}}
THE AGENT’S PLAN
{{p}}
{{jr.stamp}} {{jr.you}}

{{jr.why}}

How sure Jev was the plan fits the question{{jr.conf}}
{{jr.said}}what the report would have said
{{jr.right}}the right answer
FINAL SCORE
{{jScore}}
You
6/6
Jev, in {{jJevTime}} in total

{{jEnd}}
WHAT WE LEARNED
01
It enforces the rules it is given

Jev checks a plan against a rulebook. If the rulebook is wrong, so is the check. With a verified rulebook the agent scored 16 of 16 without Jev, which is why we would fix the rulebook first and add Jev after.

02
Its job is narrow

Counting, date math and report writing stay with the agent and the database. Jev’s only job is to say whether the plan fits the question.

03
The free option was not ready

We also tested an open-source look-alike that runs on our own machines. It was fast, but out of the box it was worse than guessing on three of seven tasks, so we parked it.

Worth adding as a fast, inexpensive check on the agent, once the rulebook is right.
KNOWLEDGE GRAPH · CASE STUDY

One graph across the warehouse and the papers.

The client’s structured data sat in the warehouse, while most of what its scientists knew about a disease, drug or biomarker sat in internal publications and reports, with no link between them. We built a knowledge graph across both, so a question about any one of those things leads to everything connected to it.

DOMAINKnowledge graph · discovery
OUR ROLEDesign and build
SOURCESData warehouse tables and internal publications
STATUSIn production
~30
Warehouse tables mapped into the graph
~1,500
Internal publications and reports linked in
~10K
Entities such as diseases, drugs, biomarkers and trials
4
Standard vocabularies used to line up names: ICD-10, SNOMED CT, LOINC, RxNorm

FIGURES ARE APPROXIMATE AND ROUNDED. EXACT COUNTS AND NAMES ARE WITHHELD TO PROTECT THE CLIENT.

TRY IT · FOLLOW THE LINKS

A simplified picture of how it connects. Click any node, or start from a question. Copper nodes are entities, gold are warehouse tables, brown are publications.

{{gsel.kind}}
{{gsel.label}}

{{gsel.note}}

DIRECTLY LINKED TO {{gsel.links}}
HOW IT WAS BUILT
01
Start from the questions

We began with the questions people bring to the data, then defined a small set of entity types to answer them: diseases, biomarkers, therapies, trials.

02
Link the sources

Warehouse columns are mapped to those entities, and the same entities are pulled out of the publications. Names are matched through standard vocabularies, so different spellings of one thing land on one node.

03
Record where every link came from

Each link stores its source, whether a table and column or a page in a document, so any answer can be traced and checked.

Answering that kind of question used to take three people and a dozen files. Now it takes one lookup, with the sources attached.
{{csCount}}
{{c.tag}}
{{c.name}}
{{c.area}}

What we solve in healthcare and life sciences.

We help executives and their teams decide faster and with more confidence. Most of the work comes down to connecting what the organization knows, governing its data, and making every piece of evidence traceable.

WHAT IT DOES

{{item.desc}}
Have a different problem? Book a data therapy consult

Sometimes the answer is something new.

Sometimes we join an organization to build something that does not exist yet, whether a business, an internal capability, or an early bet it wants to test.

NEW LINE OF BUSINESS

Start something adjacent

We take a new business from thesis to first revenue inside the enterprise, covering venture design, the product, pilot contracts, and the operating model to run it.

LABS FUNCTION

Stand up an AI lab

A separate team with its own charter and governance, set up to launch AI initiatives without getting in the way of the core business.

NEW TECHNOLOGY

Experiment at speed

Sprints of two to six weeks that put a new technology against a real business question. At the end you have a working prototype and a plan for scaling it.

We build with

We have no platform to sell. We work in the stack you already have and connect the parts that should have been talking to each other all along.

We also build, operate, and advise new ventures.

Alongside client work, we help start and grow new companies in healthcare and life sciences.

BUILDProduct, platform, and data foundations, from zero to first customers.
OPERATEFounding and fractional leadership through launch and growth.
ADVISEStrategy, go-to-market, and technical diligence for founders and boards.
{{vCount}}
LAST UPDATED 29 SEP 2026

What we are paying attention to right now.

What we are testing, and where we think it will matter in regulated industries.

{{nCount}}

How can we help?

Book a data therapy consult