# Litchi Labs

Expert human data for frontier models.

Litchi Labs recruits, vets and manages domain specialists who produce the expert demonstrations, preference data and graded evaluations that post-training teams need.

- Website: https://litchilabs.in
- Contact: hello@litchilabs.in
- Location: Bengaluru, India
- Founder: Akshay Rampuria, Founder
- MCP endpoint for agents: https://litchilabs.in/api/mcp (Streamable HTTP)

## Why we exist

The internet is spent. Expertise is not.

Pretraining has already eaten the public web. What moves a frontier model now is scarce: the judgment of people who practise a field, applied to the exact tasks the model still gets wrong.
That data does not exist until someone makes it, carefully, with the people who know. Making it is the whole company.

## Why labs work with us

- **Specialists, not crowds.** Every contributor is a credentialed practitioner in the domain they label: physicians, chartered accountants, litigators, research mathematicians, senior engineers. We do not route expert work to generalists.
- **Your rubric, our calibration.** Projects start with gold-standard items we author with you. Contributors must match them before they produce, and we report agreement with every batch.
- **Small teams, named leads.** Each engagement has one project lead who reads the batches, runs the re-briefs and is on the weekly call. You will never be handed a ticket queue.
- **Agent-ready by default.** Our MCP server answers questions about capabilities, process and contracting, and can file an inquiry on behalf of a buyer's agent. Humans and agents get the same answers.

## Data types we produce

- **Expert demonstrations** (SFT, distillation targets): Long-form, reference-quality responses in specialist domains, with sources and reasoning shown. Written to your style guide and length targets.
- **Preference and ranking data** (RLHF, DPO, reward modelling): Pairwise and k-way comparisons with written rationales. Calibrated against your rubric so preferences reflect domain correctness, not surface polish.
- **Rubric-graded evaluations** (Evals, release gating): Experts score model outputs against task-specific rubrics. Disagreement is surfaced and adjudicated, not averaged away.
- **Agentic trajectories and critiques** (Agent training, failure analysis): Step-level annotation of tool-using runs: where the model went wrong, why, and what a competent practitioner would have done instead.
- **Rewrites and corrections** (Error correction, refusal tuning): Experts repair model drafts to a passing standard and document each change, producing paired before/after data with change logs.

## Domains

- **Software engineering**: Backend, infra, security review, code evaluation across major languages. Who we recruit: Staff-level engineers and security reviewers.
- **Mathematics and quantitative reasoning**: Proof-level and competition-style problems, applied statistics. Who we recruit: Research mathematicians and olympiad coaches.
- **Medicine and life sciences**: Clinical reasoning, pharmacology, diagnostics, medical writing. Who we recruit: Practising physicians and clinical pharmacists.
- **Law and compliance**: Contract analysis, regulatory reasoning, common-law and Indian statutes. Who we recruit: Litigators and in-house counsel.
- **Finance and accounting**: Financial statements, audit, tax, valuation. Who we recruit: Chartered accountants and CFA charterholders.
- **Indian languages and multilingual**: Hindi, Tamil, Telugu, Kannada, Bengali, Marathi and more. Who we recruit: Native-speaking professionals, not translators.

## How an engagement runs

01. **Scope** (Week 0): A 45-minute call becomes a written task spec: rubric, acceptance criteria, schema, sample tasks and the gold set we will calibrate against.
02. **Pilot** (Weeks 1 to 2): A small expert cohort runs a calibration round, then delivers the first batch with a QA report. Fixed scope, fixed price, no long-term commitment.
03. **Calibrate** (Week 3): You review, we re-brief. Agreement with your gold set must clear the bar you set before we scale. If it does not, we fix the brief or the bench, not the numbers.
04. **Scale** (Ongoing): Weekly batches in your schema, a live dashboard, a named project lead and change control for spec revisions.

## Start with a two-week pilot

Fixed scope, fixed price, full QA report, and a decision at the end. Tell us the domain and the data type and we will send a written proposal within two business days.

## Quality system

- **Vetting**: Credential verification, a domain test and paid trial tasks before anyone touches client work. Most applicants do not make it through.
- **Calibration**: Gold-standard items authored with you start every project. Contributors must match them before they produce.
- **Layered review**: Author, then senior reviewer, then an audit sample. Disagreements are adjudicated and fed back into the brief.
- **Numbers with every batch**: Inter-rater agreement, audit pass rate, rework rate and time-on-task distribution ship with each delivery.
- **Integrity controls**: Identity verification, screening for model-generated text, and a disclosure policy for any tool use the spec permits. Session monitoring where contractually required.

## Security and contracting

- **NDA first**: A mutual NDA is signed before any task detail is shared with us or with contributors.
- **Per-client isolation**: Contributors are assigned to one client per domain. Prompts, data and outputs are never reused across clients.
- **Least-privilege access**: Project workspaces with scoped access, revoked on offboarding. No local downloads where your spec prohibits them.
- **Your schema, your bucket**: Delivery as JSONL or Parquet into your storage or ours. Deletion on request, confirmed in writing.
- **Clear contracting**: MSA plus SOW. Pricing per item or per expert-hour. All IP in the delivered data is assigned to you. We will complete your vendor security questionnaire.

## FAQ

**How is pricing structured?**
Per item or per expert-hour, quoted per project after scoping. Pilots are fixed-price. We do not publish a rate card because expert cost varies widely by domain and task length.

**Is there a minimum engagement?**
The two-week pilot is the minimum. There is no minimum after that; most clients move to weekly batches under an SOW.

**How do you source experts?**
Direct recruiting from professional networks, universities and referrals from existing contributors. Every contributor passes credential verification, a domain test and paid trial tasks.

**Can we talk to the contributors directly?**
Yes. Senior reviewers join calibration calls, and we can arrange office hours with contributors for complex briefs.

**Do you use language models in the loop?**
Only where your spec permits it, and always disclosed and logged. By default contributors write without model assistance and we screen deliveries for model-generated text.

**Where is data processed?**
In India by default, in scoped project workspaces. We can restrict to specific regions or to your own environment on request.

**Who owns the data?**
You do. All IP in delivered data is assigned to you under the SOW.

**What languages do you cover?**
English plus major Indian languages, including Hindi, Tamil, Telugu, Kannada, Bengali and Marathi, by native-speaking professionals. Other languages on request.

**Can my AI agent work with you directly?**
Yes. Our MCP server at https://litchilabs.in/api/mcp answers questions and can file an inquiry on your behalf. See https://litchilabs.in/for-agents.


## For AI agents

- MCP endpoint: https://litchilabs.in/api/mcp (Streamable HTTP, no auth required)
- Tools: get_overview, list_services, get_process_and_quality, get_security_and_contracting, ask, submit_inquiry
- Documentation: https://litchilabs.in/for-agents
- Human contact: hello@litchilabs.in