Applied ML
Computer vision
Detection, tracking and camera calibration for systems that have to be right on live footage.
- Object tracking
- Camera calibration
- Model training
We get AI into production
Swipe →
Automation
AI agents
Agents that do real work inside your systems, with a human gate on anything consequential.
- Tool use
- Approval gates
- Spend caps
Most enterprise AI never leaves the pilot stage. We take one workflow that matters, run it into production, and build in the spend caps and approval gates that keep it there.
Contact us- 450,000+people use software we’ve built
- 2 weeksfrom first call to a written verdict
Autonomous agent
AskCodi
Agent orchestration
MakersClaw
Multi-tenant gateway
Assistiv SDK
Legal due diligence
PaperSafe
Client site
Swanlac
Applied vision
Ball tracking
// Proof of work
Built, shipped, still running
Every product here is ours. We built them, we run them, and we pay the inference bill each month. That is why we can tell you what production costs instead of guessing.
- 450,000+
- people use software we have built
- 4
- products in the open
- 2 weeks
- to a written verdict
- 5 yrs
- running AI in production
- Our productDeveloper tools
AskCodi
An autonomous coding agent used by more than 450,000 developers. Idea to merged pull request, under a hard spend cap.
What we built
- A supervised pilot runs before anything executes unattended
- A per-run spend ceiling that halts the job, not the budget
- Every change arrives as a pull request in a repository you own
Stack
Next.js · TypeScript · Vercel · GitHub API
- Our productAgent orchestration
MakersClaw
Specialist agents working from one shared company context, wired across 16 integrations.
What we built
- An orchestrator that routes one brief across specialist agents
- Persistent company context, so each run builds on the last
- Connectors across 16 platforms, with scheduled runs in the background
Stack
Next.js · TypeScript · Vercel
- Our productLegal AI
PaperSafe
Due diligence over legal documents at volume, with a person reviewing what it finds.
What we built
- Extraction that turns PDFs, scans and Word files into structured records
- Retrieval that grounds every answer in the source passage
- A reviewer checks each extracted field before it is exported
Stack
Next.js · TypeScript · Vercel · RAG
- Client workApplied vision
Ball tracking
Ball tracking over a projected putting surface, on a detector we trained ourselves and kept fast enough for CPU.
What we built
- Detection model trained from scratch on footage we collected ourselves
- Runs on CPU, so a site needs no GPU in the rig
- Camera calibration that holds the overlay on the surface frame to frame
Stack
Self-trained model · CPU inference · Human-checked labels
// What we do
Where we take work
Five domains, one method. We pick the workflow that carries the most weight, put it into production behind cost ceilings and approval gates, then keep it running. All of it delivered end to end, not handed over as advice.
Computer vision
- Object tracking
- Camera calibration
- Model training
- CPU inference
Tracking and calibration for live footage, where the output is either geometrically correct or visibly wrong. We have trained detectors from our own collected data and kept them inside a CPU budget, which is usually the constraint that decides whether a rig ships.
Talk about this about Computer visionAI infrastructure
- Model routing
- Scoped API keys
- MCP tools
- Usage metering
- Observability
One API across providers, with scoped keys per team or per end user, tool execution over MCP, and every token and tool call metered against a wallet you can read. We built this for ourselves first. Production GenAI routinely lands at three to five times its projected cost, and nobody can fix a bill they cannot see.
Talk about this about AI infrastructureAI agents
- Tool use
- Approval gates
- Spend caps
- Audit trails
- Memory
Agents that do real work inside your systems. Anything consequential waits on a human, every run leaves a trail you can read afterwards, and a hard cap stops the job before it stops your budget.
Talk about this about AI agentsProduct builds
- Interface
- API design
- Auth and billing
- Deploy pipeline
Once a capability works, somebody still has to build the product around it. We do the interface, the backend, the billing and the pipeline that ships it, so the model does not stay stranded behind an internal script.
Talk about this about Product buildsProduction audits
- Readiness review
- Cost model
- Architecture
- Written verdict
Two weeks, fixed fee. We read a stalled build, price what finishing it takes, and name what breaks first. Sometimes the answer is that it is not worth finishing, and you get that in writing too.
Talk about this about Production audits
// The gap
Why AI stalls, and what we do about it
None of these numbers are about model quality. They are all about what happens after the demo. We built the answers before we sold them, because we run the products underneath.
88%
of AI pilots never reach production.
We take one workflow the whole way live, with a number attached to it.
46.9%
of companies are already over budget on AI.
Every token and tool call settles against one wallet, behind a cap that halts the run.
Assistiv SDK
85%
of IT leaders have had a project stopped because nobody could trace what the AI did.
Every change lands as a pull request in a repository you own.
AskCodi
40%
of agent projects Gartner expects to be cancelled by 2027, mostly on weak controls.
Nothing runs unattended until a supervised pilot passes, and a human approves anything consequential.
AskCodi
// Before you ask
The questions we get on every first call
Who actually does the work?
The people you meet on the first call. They scope it, they write it, and they are still there at go-live. Where a programme genuinely needs a hundred bodies on site we will say so rather than take the work and staff it with strangers.
What if the audit says the project is not worth finishing?
Then that is what the verdict says, in writing, and you have saved the cost of finding out the expensive way. It has happened before. We would rather lose a build than sell one we expect to fail.
Why do you charge for the first engagement?
A paid audit means we spend two weeks reading your code, your data and your logs rather than two weeks selling. It also ends the conversations that were never going anywhere, which is how we keep the team senior instead of hiring a sales floor.
Do we own what you build?
Yes. Work lands in repositories you own and runs on infrastructure you control. There is no runtime you have to keep licensing from us to keep the system alive, and no lock-in clause that makes leaving expensive.
What does it cost to run once it is live?
Maintenance typically runs 15 to 30 percent of build cost a year, and true total cost tends to reach roughly twice the initial quote across the first 18 months once data preparation, compute and drift are counted. Those numbers go in the proposal rather than surfacing in month nine.
How do you handle our data and access?
Scoped credentials, least privilege, and an audit trail of every action the system took. Models run against your data under your own provider accounts where you want that, so nothing is copied into infrastructure you cannot inspect.
How quickly can you start?
Audits usually begin within two weeks. Sprints depend on what else we have running, and you will get a real date rather than an optimistic one.
Something not covered? Ask us directly.
// Start here
Tell us what stalled
Send us what you are working on and where it got stuck. If a production audit is the right next step we will say so, and if something simpler will do we will say that instead.
- A written verdict, not a deck
- We will say if it is not worth finishing
- No retainer to get an answer
Prefer email? hello@assistiv.ai