SFT · expert generation
Demonstrations
Worked examples from people who do the job: clinical reasoning, contract analysis, financial modelling, production code.
Frontier AI
Between training a model and shipping an agent.
Two kinds of people are scarce right now. Specialists who can demonstrate, grade and break a model in their own field. And engineers who will sit inside your team and get the thing into production. We supply both, as one deployed unit.
✓ Vetted, calibrated pods
✓ Forward-deployed engineers
✓ Agreement rates you can audit
Training · evals · forward-deployed engineering
Where the work sits
The middle of the stack is where projects stall.
Labs and enterprises are well served at both ends. There are vendors who will label at volume, and there are consultancies who will write you a strategy. The gap is the part where a model has to become something that works on a Tuesday, in your systems, against your data.
Upstream
Training the model
Expert demonstrations, preference ranking, evaluations and red-teaming, produced by people who hold the credential the task needs.
Where we sit
Getting it to work in your context
Forward-deployed engineers embedded in your team, closing the distance between a model that scores well and a system your business actually runs on.
Downstream
Agentic development
The agents, tools and evaluation harnesses that turn a capability into a product. We build these for ourselves, which is why the engineers are worth deploying.
What we produce
The work your team keeps queuing for.
Commodity labelling stopped being the bottleneck. What is scarce is a specialist who understands the task well enough to demonstrate it, grade it, or break it.
Preference · RLHF
Ranking and preference
Pairwise judgments against your rubric. Inter-rater agreement gets measured and reported with every batch.
Evals
Evaluation and grading
Rubric design, benchmark grading and regression suites, built by people who run eval harnesses on their own systems.
Safety
Red-teaming
Adversarial probing and jailbreak discovery, written up so your safety team can reproduce and act on every finding.
Multilingual
Urdu and regional languages
Native fluency in Urdu, Punjabi, Sindhi and Pashto alongside professional English — coverage most vendors quietly outsource.
Agents
Agent evaluation
Task-based testing in real environments — browsers, codebases, workflows — scored against criteria you define up front.
How you get it
One team that stays on your account.
You brief a pod once. It keeps the context, the corrections and the standard, and it is still there next quarter.
- 01
We source for your domain
Graduates and practising professionals out of LUMS, NUST, IBA, FAST and GIKI, plus licensed specialists where the work needs a credential.
- 02
They get screened before they touch your systems
Language assessment, domain testing and a reasoning screen. We publish the pass rates so you can see what the filter actually removes.
- 03
The pod calibrates to your rubric
They train against your guidelines until agreement clears the threshold you set, and we absorb the cost of getting them there.
- 04
Work ships with its own evidence
Every batch arrives with sampling results, double-marked disagreements and the agreement rate attached.
- 05
The same people stay on your account
Context compounds instead of resetting. Your rubric does not get re-explained to a new stranger every week.
Forward-deployed engineering
An engineer inside your team, not a ticket queue.
An FDE sits with the people who own the problem, writes production code against your stack, and stays long enough to be held to the outcome. Ours come off systems we operate every day, so the first week is spent on your problem rather than on learning what an agent is.
- 01
They sit where the problem is
In your standups, your repo and your incident channel. The distance between the person who understands the workflow and the person writing the code goes to zero.
- 02
They write production code
Agents, tools, retrieval, evaluation harnesses and the plumbing between them. Reviewed by your engineers, merged into your repositories, owned by you.
- 03
They build the eval before the demo
A capability nobody can measure is a capability nobody will sign off. The harness that proves the thing works gets built alongside the thing itself.
- 04
They hand it over and document it
The engagement ends with your team running the system without us. Anything else is a dependency we sold you, and it would show up in the second invoice instead of the first.
Where the supply comes from
Pakistan is the talent pool the industry has not priced in.
The same labs paying US rates for expert hours are competing for people who work remotely anyway. Pakistan already exports this work through freelance platforms at scale. We package it with vetting, calibration and accountability.
$4.5BPakistan IT exports, FY2026, up 29% year on year
$950MFreelance earnings in ten months
Top 5Country on global freelance platforms
5Universities feeding the vetting funnel
Built to pass your vendor review.
Procurement asks the same questions every time. Here are the answers before you ask: every contributor signs an NDA before their first task, access is least-privilege and revoked on rotation, and work runs inside your stack when your policy requires it.
SOC 2 Type II and ISO 27001 are still on the roadmap. Until they land we start you on non-sensitive and public-data pilots, and we say so up front rather than letting you find out in diligence.
NDA before first task
Signed by every contributor. Client work is never reused, resold or shown in a portfolio.
Least-privilege access
Scoped credentials, revoked on rotation, with an audit trail per batch.
Measured QA
Sampling, double-marking and agreement rates reported with the work.
Pilot first
A bounded batch before any commitment. Judge the output, then scale the pod.
Your tools
We work inside your annotation stack, so nothing migrates and no data leaves your perimeter.
Named supervision
A pod lead who answers for quality, reachable in your timezone overlap.
How this is bought today
Three ways to buy expert work, and what each one costs you.
Most teams have already tried at least one of these. The differences show up in rework, not in the first invoice.
| Crowd marketplace | Traditional BPO | Kitsune pod | |
|---|---|---|---|
| Who does the task | Whoever claims it that hour | Whoever is on shift | A named pod that stays on your account |
| Domain credential | Self-declared | General graduate pool | Tested, and licensed where the work needs it |
| Calibration to your rubric | You write guidelines and hope | Billed to you as training hours | Done before billing starts, at our cost |
| Quality evidence | Spot checks you run yourself | A throughput report | Agreement rate and double-marked disagreements per batch |
| Context between batches | Resets every time | Resets on rotation | Same people, corrections compound |
| Who answers when quality slips | Platform support | An account manager | The pod lead who signed the batch |
| Language coverage | English-first | English-first | Professional English plus Urdu, Punjabi, Sindhi, Pashto |
What a first engagement looks like
Two weeks, one bounded batch, no commitment attached.
Days 1–3
Scope and rubric
You send the task, the volume and the standard. We come back with the sample design, the pass criteria and what the batch will cost, in writing.
Days 4–8
Pod screening and calibration
Contributors are screened for your domain and trained against your guidelines until agreement clears the threshold. Those hours are ours, not yours.
Days 9–14
Batch and evidence
Work lands with sampling results, the agreement rate and every double-marked disagreement attached, so your reviewers can audit it instead of trusting it.
Common questions
Answered before procurement asks.
What is the smallest engagement you take?
A single bounded pilot batch. We would rather you judge a small piece of real output than sign a volume commitment against a capability deck. If the pilot does not clear your bar, that is the end of it and you keep the work.
How do you screen contributors?
Language assessment, a domain test written against the task you are actually buying, and a reasoning screen. Where the work needs a licence — clinical, legal, financial — we verify the credential instead of taking a claim on a CV.
How is quality measured rather than asserted?
A sample of every batch is double-marked by a second qualified contributor. We report the inter-rater agreement rate and hand over the disagreements themselves, so you can see where the rubric is ambiguous instead of only where the workers were wrong.
Do you have SOC 2 or ISO 27001?
Not yet. Both are on the roadmap and neither has landed. Until they do we start clients on non-sensitive or public-data work, and we say so here rather than letting it surface halfway through your vendor review.
Can the work run inside our own tooling?
Yes, and it usually should. We work inside your annotation stack under scoped credentials, so your data stays in your perimeter and nothing has to be migrated to us and migrated back.
Who owns what the pod produces?
You do, on delivery. Client work is never reused for another engagement, resold, or shown in a portfolio, and every contributor signs an NDA before their first task.
Why Pakistan?
The country already exports this class of work at scale through freelance platforms, at rates well under US and European equivalents, with a large English-fluent graduate pipeline out of LUMS, NUST, IBA, FAST and GIKI. What has been missing is a vendor wrapping that supply in vetting, calibration and accountability. That is the whole business.
What timezone overlap do we get?
Pakistan Standard Time is UTC+5, which gives you a full working overlap with Europe, the Gulf and India, and a morning overlap with the US east coast. Pod leads are reachable inside your working hours, not at the end of a queue.
How fast can a pod scale?
A calibrated pod grows faster than a cold one because the standard is already documented and the existing members train the new ones. Realistically, doubling a pod takes two to three weeks including screening and calibration, and we will tell you if your timeline needs more than that.
Start small
Send one batch. Judge the output.
Tell us the task, the volume and the standard. We will scope a bounded pilot, run it with a calibrated pod, and hand it back with the quality evidence attached.