APractice 01
Direct to the model.
We integrate against the LLM API the same way we integrate against Postgres. No agent frameworks, no orchestration vendors, no migrations every six months.
0Frameworks added to your stack
NewAI readiness · Development · Design
AI products,built onsolid ground.
Kensink is an AI readiness, development and design house. Senior engineers and designers audit your foundations, design the experience and ship production AI in eight weeks. No framework lock-in, no orchestration vendors, no slide decks.
- Engagements shipped to production
- 47
- Framework licenses, ever
- $0
- Median kickoff to production
- 8 wk
- Availability
- 2 build slots openLead time 3–4 weeks
Click to pulse the coreScroll
- TypeScript
- Python
- Postgres
- pgvector
- Anthropic
- OpenAI
- Redis
- ClickHouse
- vLLM
- OpenTelemetry
- Next.js
- Braintrust
- Grafana
Shipping production AI for
- NuforceSeries-B insurer
- Affidavit MappCapital markets
- AICoachVertical SaaS
- ConfidentialB2B marketplace
- ConfidentialEHR add-on
01Our point of view
Our point of view
Most companies aren't short on AI ideas. They're short on solid ground: data they can trust, infrastructure that ships, evals that catch failure and interfaces people believe. We build that ground first, then the product on top of it.
01
Readiness
Know what your foundations can carry before you spend on a build.
02
Development
Senior engineers ship production systems directly against model APIs.
03
Design
Designers shape trust, fallbacks and feedback inside the same squad.
02What we do
One squad for readiness,development and design.
Strategy decks don't ship. We go from an honest read of your foundations to a designed, tested system in production, with the same senior people the whole way.
012 weeks · fixed fee
AI Readiness Audit
Two weeks to know exactly where you stand. We score your data, infrastructure, governance and team, then hand you a ranked use-case map and a build plan.
- Foundations scorecard
- Ranked use-case map
- Build plan and budget
02Research · UX · UI
AI Product Design
Interfaces people actually trust. Conversational flows, copilots, confidence cues and graceful failure states, designed with your users and shipped in code.
- Trust patterns
- Failure states
- Design systems
036–12 week builds
Agents & LLM Apps
Production systems that call models directly. Triage, routing, research and generation on your stack, on your data, behind your auth.
- Triage & routing
- Copilots
- Generation
04Evals · red team · guardrails
Evals & AI Security
Behavior tests, red-teaming and guardrails, so you know what your system does before your customers find out.
- Behavior suites
- Red-teaming
- Guardrails
05Web · iOS · Android
MVP & Mobile
From prototype to app store. Web and mobile products with AI at the core, built by the same senior squad.
- Web apps
- Mobile apps
- AI at the core
06CCPA · US/EU residency
Enterprise Programs
Multi-phase roadmaps for larger organizations, with data residency in the US or EU and compliance designed in from day one.
- Phased roadmaps
- Data residency
- Compliance by design
03AI readiness check
How AI-ready are you,really?
Six questions, about ninety seconds. Score your foundations the way we do in week one of every engagement, and see which kind of engagement fits where you are today.
0 / 6 answered
Readiness score
--/100
04How we work
Four things holdevery build up.
Pick the practice closest to your blocker. They are the four rules we don't bend, whichever shape the engagement takes.
BPractice 02
Senior, all the way down.
Every line is written by an engineer who has shipped before. There's no junior team behind the curtain, no handoff to a pyramid and no surprise subcontractors.
12 / 0Senior engineers / juniors
CPractice 03
Designed, not decorated.
Designers sit inside the build squad. Confidence cues, fallbacks and feedback loops are designed alongside the prompts, so people trust what ships.
1Squad for design and engineering
DPractice 04
You own the code.
The repo, the prompts, the evals and the runbook. Your team runs it at handoff, MIT licensed, with no vendor lock-in and no black boxes.
100%Of the code is yours
05From production, right now
A live system,week seven of eight.
Claims triage for a Series-B insurer. Direct LLM integration over a Postgres claims pipeline. Every first notice of loss is read, classified and routed in under half a second.
Liveclaims-triage/prod/us-east
Build87.5%
- Auto-resolved
- 73.0%
- p50 latency
- 412ms
- Eval pass rate
- 98.2%
Claims resolved per minute80
Auto-resolved73%
Human review21%
Fraud flag6%
Decision streamp50 · 412ms
- CLM-48213382msWater damage·Auto-approved·conf 0.96
- CLM-48214417msAuto collision·Routed to adjuster·policy check
- CLM-48215391msStorm·Auto-approved·conf 0.93
- CLM-48216446msTheft·Fraud signal·human review
- CLM-48217356msGlass·Auto-approved·conf 0.98
- CLM-48218431msLiability·Escalated·conf 0.58
- Cost on human-handled tasks
- −80%
- Throughput by day 90
- +60%
- Frameworks shipped
- 0
06Engagement shapes
What we ship ineight weeks.
Three engagement shapes, each with a fixed scope and a fixed fee. You can see everything we leave behind before we start.
01 · Operational01
Triage & deflection systems
Tickets, claims and support emails. We sort, summarize and auto-resolve the long tail, and humans keep the hard cases.
- Stack
- Postgres · Python
- Build
- 6–8 weeks
- Typical lift
- 40–75% deflection
Fixed scope · fixed fee
02 · Internal02
Research & analyst copilots
Your knowledge base, filings and contracts, wired into the tools your team already opens at 9 a.m.
- Stack
- pgvector · Next
- Build
- 8–10 weeks
- Typical lift
- 3–5 hrs / analyst / wk
Fixed scope · fixed fee
03 · Product03
Customer-facing AI features
A model-powered feature inside your product, shipped with the eval suite, kill switch and rollout plan your team can run.
- Stack
- TS · pgvector
- Build
- 8–12 weeks
- Typical lift
- +18% activation
Fixed scope · fixed fee
Not sure which shape fits? Score your foundations first, it takes about ninety seconds.
Take the readiness check07Results, not decks
Numbers fromforty-sevenshipped engagements.
- Median time from kickoff to a build serving production traffic.
- 8wk
- Average cost reduction on the human tasks we automate.
- −80%
- Mean productivity lift, measured at day 90 after launch.
- +60%
- Average first-year ROI, weighted by engagement size.
- 340%
08The build
A four-stage build,no theatre.
Each stage has one deliverable and one decision point. No status meetings, no slide decks. Just the next thing, on a Tuesday.
Week 1Discovery
Scope memo
Two senior architects on site for a week. Out comes a written brief naming the smallest valuable thing to ship, and how it should feel to use.
DecisionWhat we build, and what we don't.
- On-site
- Experience map
- Memo
Weeks 2–6Build
Production system
Code, prompts, evals, infrastructure and interface. Direct LLM integration on your stack, on your data, behind your auth.
DecisionShip to a pilot group.
- pgvector
- Python
- TypeScript
Week 7Eval
Behavior tests
A test suite for behavior, not just text, plus dashboards your team will actually open on a Monday morning.
DecisionGo live, or tune.
- Braintrust
- Grafana
Week 8Handoff
Runbook + pairing
A week of pairing with your engineers, a runbook they can follow at 2 a.m. and a 90-day warranty on everything we built.
DecisionPhase two, or done.
- Pairing
- 90-day warranty
09Selected work
Recentproductionsystems.
Five of the last twelve engagements. Each one is running today, measured against the number we agreed on in week one.
- Series-B insurerApr 2026
73%
Auto-resolved
- The problem
- Every incoming claim email was read by a person before it could be routed, and the queue grew with every storm season.
- What we shipped
- A direct-to-model triage engine on the existing claims database, with evals per claim type and a human-review lane for edge cases.
- The outcome
- 73% of claims now resolve automatically, at 80% lower cost on the work people used to do by hand.
- Stack
- PostgresPythonEvals
- Capital markets · 200 analystsMar 2026
4 hrs
Saved per analyst, weekly
- The problem
- Analysts spent their mornings stitching filings, transcripts and internal notes together before they could form a view.
- What we shipped
- A research copilot over filings and internal memos that answers with citations, inside the tools analysts already use.
- The outcome
- Roughly four hours back per analyst every week, across a team of two hundred.
- Stack
- pgvectorNext.jsCitations
- B2B marketplaceFeb 2026
+4.2 pt
Margin
- The problem
- Sellers priced long-tail listings by gut feel and left margin on the table.
- What we shipped
- A recommendation layer that proposes prices with reasons, and an approval flow sellers were happy to use.
- The outcome
- A 4.2 point margin lift on recommended listings.
- Stack
- ClickHousePythonTypeScript
- Vertical SaaSJan 2026
41%
Tickets deflected
- The problem
- A small support team answered the same setup questions hundreds of times a week.
- What we shipped
- An answer layer in front of the help desk that resolves known issues and hands the rest to a person with full context attached.
- The outcome
- 41% of tickets are resolved before they reach a person.
- Stack
- pgvectorRedisTypeScript
- EHR add-onNov 2025
FDA
Cleared path
- The problem
- Clinicians needed concise visit summaries without unreviewed AI output landing in the record.
- What we shipped
- A summarization tool with clinician sign-off built into the flow, full audit trails and a documented regulatory path.
- The outcome
- On an FDA-cleared path, with every summary reviewed before it reaches the chart.
- Stack
- PythonPostgresAudit trail
10Product-ready builds
Some of our buildsyou can just buy.
Most of our work is bespoke. A few builds hardened into products you can roll out today, each one a real system running in production. Seven are shipping now, here are four.
Enterprise AILive
OU.Chat
Your enterprise AI. Your data. Your control.
Customer experienceLive
Omazy CX
One platform for every customer conversation.
Finance operationsLive
NuForce AI
Your finance back office, run by AI you own.
VoiceLive
AI Receptionist
A front desk that never misses a call.
11Straight answers
What CTOsactuallyask us.
The short versions of the questions that come up on every first call. If yours isn't here, ask it directly.
Yes, forty-seven times so far. The trick is that we scope the smallest valuable thing in week one, in writing, and we hold that line. If something bigger is worth doing, it becomes phase two instead of scope creep.
A two-week, fixed-fee look at your data, infrastructure, governance and team. It ends with a scorecard, a ranked use-case map and a build plan you can run with us or without us.
Then we split it. Larger work becomes a multi-phase program or an ongoing partnership, and every phase still ends with something running in production.
Less surface area, fewer abstractions to debug at 2 a.m. and smaller bills. We call model APIs directly, the way you'd talk to Postgres, so your team can read every line.
Yes. Designers sit in the same squad as the engineers, so the interface, the trust cues and the failure states are designed alongside the model calls rather than painted on afterwards.
From day one. The repository, prompts, evals and runbook live in your accounts. No license fees and no black boxes.
Two named senior engineers you meet in week one. No juniors, no handoff to an offshore pyramid and no surprise subcontractors.
Every build ships with a 90-day warranty. If it breaks in the way we built it, we fix it.
12Start a project
Bring a real problem. We'll bringcode on day one.
No discovery fees and no statement of work that takes a month to sign. Tell us the problem and a senior engineer reads it, not a sales team.
047
Engagements
8 wks
To production
100%
Code you own
