Shiv Shankar
Founder & CEO
Founder and CEO of GONATURELLE PRIVATE LIMITED. Sets the company's direction and leads customer, partner and investor relationships from the Chennai headquarters.
We test the agents that cannot afford to be wrong.
GONATURELLE grew out of a one-line prompt edit that silently broke a medical-necessity agent. We build the harness that would have caught it.
- legal_name
- GONATURELLE PRIVATE LIMITED
- incorporated
- Chennai, India · 2026
- cin
- U47912TN2026PTC190616
- headquarters
- Mannady, Chennai, IN
- data_regions
- us-east-1 · ap-south-1 · your VPC
- attestations
- SOC 2 Type II · HIPAA BAA
- trains_on_customer_data
- false
sha2564be1c07f93d2a8e65f10c4b7a29de3086f5b1c72
The bug that passed every test.
In late 2024 the engineers who would become GONATURELLE's founding team were working on a payer prior-authorization product. An engineer shortened the system prompt to cut token cost. Final-answer accuracy on the team’s test set did not move. In production, the agent started pulling benefits before verifying the member and, on some imaging requests, skipped the coverage-criteria check entirely. A reviewer caught it eleven days later.
Rahul Venkataraman, now CTO, had seen the same pattern in lending: a credit-servicing agent whose answers looked right while its tool calls drifted. Neither team lacked tests. They lacked tests of the trajectory, judged by people who knew what a correct clinical or credit decision looks like, wired into the pull request instead of a quarterly review.
Generic benchmarks such as MMLU say little about whether an agent respects a formulary, a Regulation B reason code or a policy exclusion. Through 2025 the founding team built the first physician-labelled golden dataset with Dr. Meera Iyer. In 2026 Shiv Shankar founded GONATURELLE PRIVATE LIMITED in Chennai to turn that work into a product.
main
- verify_member()
- fetch_benefits()
- check_lcd_criteria()
- submit_prior_auth()
PR #212
- fetch_benefits()ran before identity check
- verify_member()
- submit_prior_auth()criteria check skipped
final answer 0.94 → 0.94
trajectory 2 violations
“In a regulated workflow, how the agent reached the answer is part of the answer. If your evals only read the last message, you are testing a different system from the one you ship.”
Benchmarks are not validation
Leaderboard scores measure general ability. Regulated teams need evidence on their own workflows, rubrics and edge cases.
Prompts are code
A prompt edit deserves the same regression gate as a schema migration: diffed, tested against a baseline, blocked when it breaks.
Judges need calibrating
An LLM judge is a measuring instrument. Without agreement statistics against experts, its scores are opinions.
Six operating principles.
The rules we hold ourselves to when a deadline, a customer or a metric pushes the other way.
- 01Measure agreement before trusting a scoreNo judge ships until we know its kappa against expert consensus on the exact rubric it will apply. If we cannot measure a judge, we do not use it.
- 02The gate is the productA dashboard nobody opens protects nobody. Evaluation has to run where code changes: on the pull request, fast enough that nobody turns it off.
- 03Experts label, models assistPhysicians, attorneys, CPAs and actuaries write and adjudicate the golden cases. Synthetic expansion multiplies their work; it never replaces their judgment.
- 04Customer data never trains anythingZero data retention with every model provider we call, no training on customer cases, and no prompts or PHI in our own telemetry. Not on any plan, not for any reason.
- 05Say exactly what a result meansA passing gate is evidence, not a certificate. Our reports state their scope, version and limits, and we do not let marketing round them up.
- 06Publish the methodCalibration methods, pack agreement statistics and the Reliability Index methodology are public so customers, auditors and competitors can check our work.
Leadership.
Engineers, a physician, a lawyer and a security lead. Every pack we ship has someone on this list accountable for it.
Rahul Venkataraman
CTO
Led ML platform engineering at Infosys and later a US lending fintech, where he owned the CI pipeline for credit models. IIT Bombay (CSE). Designs the trajectory assertion engine.
MI
Dr. Meera Iyer
Head of Clinical Evaluation
MBBS and MD (Internal Medicine) from AIIMS New Delhi; eight years in hospital medicine and clinical informatics. Runs the physician panels that annotate GONATURELLE's clinical golden datasets.
KR
Karthik Reddy
Head of Judge Calibration Research
PhD in machine learning from IISc Bengaluru on annotator disagreement. Owns the calibration engine, agreement statistics and the Reliability Index methodology.
NJ
Nandini Joshi
Head of Regulatory Science
LL.M. from National Law School of India University; previously model-risk and compliance counsel at a global bank's Mumbai technology centre. Maps statutes to test suites.
VC
Vikram Choudhary
VP Engineering, Eval Infrastructure
Ran distributed build and test infrastructure at Wipro and a Bengaluru developer-tools company. NIT Trichy. Owns the sandboxed runners and the sub-90-second CI gate.
IB
Ishita Banerjee
Head of Security & Compliance
Security engineering lead at Tata Consultancy Services for BFSI clients; CISSP. Runs GONATURELLE's SOC 2 Type II and HIPAA programmes and the VPC deployment hardening.
Advisory board
Practitioners who review our pack designs and tell us when a rubric would not survive a real review board.
- SRDr. Suresh RaghunathanClinical Informatics Advisor
Former chief medical information officer of a multi-hospital network in Chennai; advises on clinical safety review boards and guideline-grounded evaluation.
- LMLakshmi MenonModel Risk Advisor
Two decades in model validation at international banks, including leading SR 11-7 validation teams; advises on credit and fraud agent testing.
- AKArvind KulkarniInsurance & Actuarial Advisor
Fellow of the Institute of Actuaries of India; former head of claims analytics at a general insurer. Advises on claims adjudication and coverage packs.
Milestones, 2025 to 2026.
Every certification and release, dated. Trust items are the ones auditors ask about first.
2025
- Feb 2025Company
Founding team starts work
Engineers from clinical and lending AI teams start building trajectory tests and a physician-labelled dataset, ahead of incorporation.
- May 2025Product
First clinical dataset: prior authorization
1,120 medical-necessity cases, each labelled by two physicians from Dr. Meera Iyer's panel, with adjudicated disagreements and published agreement.
- Jul 2025Product
Trajectory assertion engine prototype
Sequence, schema and parameter-boundary assertions over captured tool calls, used first by three regulated design partners.
2026
- Jan 2026Company
GONATURELLE PRIVATE LIMITED incorporated
Shiv Shankar founds the company in Chennai (CIN U47912TN2026PTC190616) to take the founding team's work to market.
- Jan 2026Product
GitHub Action generally available
gonaturelle/gate-action@v1 posts regression diffs to pull requests and fails the check on regression. GitLab CI and CircleCI follow in March.
- Feb 2026Trust
HIPAA BAA programme on every plan
Business associate agreements offered on Developer, Scale and Enterprise, with PHI detectors on ingest.
- Mar 2026Trust
India data residency
Multi-tenant region in Mumbai with disaster recovery in Hyderabad, alongside the US region. Customer data stays in the region chosen.
- Apr 2026Trust
SOC 2 Type I report
Point-in-time audit of security, availability and confidentiality controls, issued by an independent CPA firm.
- Jun 2026Product
EU AI Act report templates
Annex IV technical documentation sections, FDA GMLP-aligned validation summaries and bank model-risk governance packs, generated from signed run manifests. As of November 2026, Annex III high-risk obligations apply from 2 December 2027.
- Jul 2026Product
gate-action@v2 with fast path
20 critical cases on every push, full suite on merge. Median fast-path gate of 74 seconds in GitHub Actions (GONATURELLE platform telemetry, Q3 2026).
- Sep 2026Research
Judge calibration study
Calibrated ensembles reach kappa 0.87 against expert consensus versus 0.42 for an uncalibrated single-prompt judge, across 6,400 double-annotated cases.
- Oct 2026Trust
SOC 2 Type II report
Six-month observation window, April to September 2026. Report and bridge letter available in the Trust Center under NDA.
- Nov 2026Research
Vertical AI Reliability Index, first release
48 models across 12 regulated task domains, scored by calibrated ensembles, free to cite under CC BY 4.0.
Corporate details.
The facts procurement and security teams ask for, in one place.
- Legal entity
- GONATURELLE PRIVATE LIMITED
- Incorporation
- Private limited company, India, 2026
- CIN
- U47912TN2026PTC190616
- Headquarters
- Chennai, Tamil Nadu, India
- Sales
- +91-44-4123-3939
- Support (India toll-free)
- 1800-266-3939
- General
- hello@gonaturelle.si
- Security
- security@gonaturelle.si
- Legal and privacy
- legal@gonaturelle.si
- SOC 2 Type II
- HIPAA BAA on every plan
- Zero data retention with model providers
- US and India data residency
GONATURELLE PRIVATE LIMITED
No. 14/11, Sunkurama Street, Mannady, Tondiarpet Fort St George, Chennai, Tamil Nadu 600001, India
No. 14/11, Sunkurama Street, Mannady, Tondiarpet Fort St George, Chennai, Tamil Nadu 600001, India
Stop testing regulated agents on generic prompts.
Run your agent against 100 industry-standard clinical or financial edge cases in 5 minutes. 14-day free production trial, no card.