OnFinance AI

The company built its own domain model and everything runs on it. NeoGPT is fine tuned on more than 300 million tokens of regulatory text for clause level reasoning. More than seventy agents sit above it, interpreting circulars, assigning owners, scoring controls, drafting filings and answering questions with citations. The chief technology officer states the premise directly, that generic AI cannot solve this sector's problems. Without the model, what remains is a task tracker with a document library, which is the manual process the platform replaces.

The agents are designed to route work to people rather than to conclude it, interpreting a circular at clause level then assigning tasks to the right teams, tracking deadlines and generating audit ready evidence, with the compliance officer remaining the accountable party throughout. Answering with citations lets a reviewer check any interpretation against the source instrument.

What is not described is the gate before an output leaves: no stated review step before a drafted filing is submitted to a regulator, no confidence indication on a clause interpretation, and no escalation path where agents disagree with a human assessment.

Provenance is strong and performance evidence is weak. The base model is named openly, an identified open weight model fine tuned for the domain, and the training corpus is quantified and attributed to named regulators. Base models are rarely named in this market. Citation backed answers make any interpretation traceable to the instrument behind it, and explainability is stated as a founding goal. Against that, no accuracy or evaluation results are published, no model documentation exists, and regulator grade outcomes are asserted without a benchmark.

The named customer list is exceptional for a company founded in 2023 and includes institutions at the centre of the market: both national stock exchanges, a major private bank, several large asset managers and a listed broker. Outcomes are quantified against a natural unit, with institutions reported to save more than a hundred hours of manual effort per regulatory circular and processes that took weeks completing in minutes.

Backing came from a leading regional venture programme with participation from an operator at a frontier model lab, and independent tracking shows headcount up 142 percent year on year. Coverage spans more than forty regulatory domains.

Three disclosures matter. The training corpus is specified by size and source, more than 300 million tokens drawn from named regulators' publications, so a buyer knows what shaped the model. Responses are described as hallucination resistant and citation backed rather than hallucination free, which is the accurate formulation. And on premise deployment means client documents do not pool anywhere central, foreclosing the cross customer question by architecture. What is absent is evaluation evidence for clause level interpretation and any account of how the regulatory corpus is refreshed as circulars issue.

The substantive privacy answer is on premise deployment. A bank's internal policies, control assessments, draft filings and audit evidence stay inside its own environment rather than traversing a vendor's cloud. The company describes its agents as secure and on premise for internal teams. That structurally forecloses most of the usual privacy questions.

Indian data protection law and central bank localization expectations apply to these institutions and are met more easily by that architecture. No privacy framework, retention schedule or subprocessor list is published.

Two named certifications are referenced, ISO 27001 and a SOC report, so the claim names standards rather than asserting unspecified certifications. Deployment inside national exchanges and a major private bank implies those assurances were tested in procurement. What is absent is the surrounding surface: no trust center, no report request path, and no stated audit period or scope.

The regulatory grounding is specific, and it is the product rather than a claim. Named regulators at home are the securities board SEBI, the Reserve Bank of India, the insurance authority IRDAI and the mutual fund association AMFI. For international work the US SEC, FINRA and a ratings agency are named, and the corpus is built from their publications.

The platform tracks consultation papers so a firm can anticipate rule changes before they take effect, which is engagement with the rulemaking pipeline rather than the rulebook alone. More than forty regulatory domains are covered.

The subjects are regulations and controls rather than people, so this reads as accuracy governance, and the stakes are concentrated: a clause interpreted wrongly propagates into task assignments, control scores and filings submitted to a regulator, and the same error repeats across every institution using that interpretation. Explainability is a stated design goal and citations support it.

What is missing is measurement of correctness: no accuracy figures for clause level interpretation, no error analysis by circular type or regulator, and no description of what happens when the model and a compliance officer read a provision differently.

Citations are the practical recourse mechanism and they are well matched to the task, because a compliance officer can verify any interpretation against the circular it came from before acting, and on premise deployment keeps the evidence trail inside the institution. Nothing binds the vendor beyond that. No accuracy guarantee despite regulator grade outcomes being claimed, no remediation term where a misread clause produces a defective filing, and no published error rate. The exposure sits entirely with the institution and the officer who signs the submission.

The full model chain is disclosed, base model included. NeoGPT is stated to be a fine tune of a specific open weight model at a stated parameter size. The training corpus is quantified at more than 300 million tokens and attributed to named regulatory publishers. On premise deployment means inference runs in the customer's environment rather than through an external provider. A buyer can therefore trace the model's origin, its training data and where it executes. In this market that chain is usually left open at one point or another.

The integration that matters most is outward to the regulators, with agents monitoring regulator portals directly and fetching circulars, master directions and gazette notifications as they publish, which is the ingestion problem this market actually has. Inside the institution, agents assign tasks across teams and generate audit ready evidence, implying workflow integration. On premise deployment shows the platform can run inside a bank's own environment. Named connectivity to internal systems is not published: no governance, risk, document or core platforms are identified, and there is no public developer documentation.

On premise deployment is offered and described as a core property rather than an option. That is the correct architecture for a market where the central bank imposes data localization and institutions are examined on where regulatory data rests. A separate server location is also referenced for international work, indicating deliberate regional separation as the company expands toward the Middle East and the United States. What is not published is the detail: no residency documentation, transfer mechanisms or subprocessor list.

The underlying model is offered alongside two distinct platforms, so the scope of a price is a live question. Deployment is on premise, which usually implies an implementation component. No rates, tiers, billing unit or minimum are published, and nothing public indicates whether pricing follows agents deployed, regulatory domains covered, users or an enterprise license.

Institution types are enumerated precisely and cover the whole regulated perimeter of its market, spanning banks, non banking financial companies, asset management companies, brokers, exchanges and insurers, with named deployments across at least four of those categories including both national exchanges. The framing is operational rather than aspirational, observing that each of those institutions runs seventeen to thirty five regulatory workflows in parallel. Functional reach extends beyond compliance into equity research and credit underwriting, and expansion is stated toward the Middle East and the United States.