What does the free audit actually involve?
+
You send us 50 model outputs or RLHF preference pairs — securely, via a signed URL to an encrypted S3 bucket. We evaluate them for sycophancy susceptibility and/or hallucination rate depending on what you need. You receive a 1-page findings report within 5 working days. No cost. No sales call required beforehand. No obligation to engage further. If the finding is not interesting, you have lost nothing. If it is, we discuss next steps.
What is the minimum project size?
+
For RLHF preference data: 500 pairs minimum. For NLP annotation: 300 documents minimum. For image annotation: 500 images minimum. For evaluation audits (sycophancy, hallucination, red-teaming): these are fixed-scope projects with no minimum volume requirement — we scope them based on your model and risk profile. We do not have a minimum spend threshold.
How do you handle data confidentiality?
+
A mutual NDA is signed before any data exchange. Your data is stored in an isolated, encrypted S3 bucket — one bucket per client, never shared. Access is restricted to named individuals listed in your project data card. Every annotator signs an individual confidentiality agreement before they see any task. We are DPDP Act 2023 compliant for Indian data, GDPR-ready for European data, and HIPAA-aligned for healthcare data. After delivery, data is deleted from our systems on your request.
How long does a typical project take?
+
After the Scope of Work is signed: scoping and guidelines take 3–5 days. Calibration takes 2–3 days. Annotation time depends on volume — typically 1,000 RLHF pairs take 5–8 working days of annotation, 5,000 NLP documents take 7–10 days. QA and delivery adds 2–3 days. Total from signed SOW to delivery: typically 2–4 weeks depending on volume and complexity. We include a project timeline in every SOW.
What quality guarantee do you offer?
+
We guarantee a Cohen's kappa inter-annotator agreement score of ≥ 0.70 on every delivery. If a batch falls below this threshold, we recalibrate and re-annotate the affected tasks at no cost. Every delivery includes a QA report with the actual kappa score, gold standard pass rates, and batch error log — so you can verify the quality claim yourself rather than taking our word for it. Two weeks post-delivery, we follow up for your model benchmark result. If the data did not produce the expected improvement, we investigate and re-deliver.
Do you work with Indian AI startups or only large enterprises?
+
Both — and our pricing is designed to work for both. Indian AI startups at Seed and Series A stage are our core ICP. We have project-based pricing starting at ₹3L that works for teams that need 500 RLHF pairs on a startup budget. We also work with MNC India AI labs and enterprise teams deploying GenAI products, where projects are larger and often convert to monthly retainers. The free audit is available to everyone regardless of company size.
Can you annotate in Indian languages?
+
Yes — this is one of our core capabilities. We maintain native-speaker annotator pools for Hindi, Tamil, Telugu, Kannada, Malayalam, Bengali, Marathi, and Gujarati. For RLHF annotation in Indian languages, our annotators are not just native speakers but also culturally aware of communication norms specific to each language — which affects how sycophancy, politeness, and directness are interpreted in preference data. We also provide cultural context notes in guidelines for cross-language annotation consistency.