VOX AI (d.b.a. Sharly.ai)

AI-powered document research platform — summarise, verify, cite, and collaborate across dense documents with source-backed precision

We backed VOX because Simone Macario had built something genuinely difficult: a conversational AI system that understood context, handled ambiguity, and navigated the messiness of real human language at scale. The F&B chatbot was the application; the capability underneath it was what we were investing in. When the team redirected that capability toward the problem of document intelligence — helping knowledge workers extract meaning from dense, complex documents without hallucinating sources — it became clear that the pivot hadn't abandoned the thesis. It had found a much larger market to apply it to.

The original bet: conversational AI before it was mainstream

When we first invested in VOX, the conversational AI landscape was meaningfully less mature than it is today. Building a system that could handle the genuine variety of human language — intent recognition, context management across a conversation thread, appropriate escalation when the machine reached its limits — required real technical depth in natural language processing. Most chatbots at the time were decision-tree systems masquerading as AI: ask them something outside the script and the experience fell apart immediately.

VOX, under Simone Macario's leadership, built differently. With a background in computer engineering and experience across Italy, China, and Singapore before founding VOX in Malaysia, Simone brought a linguist-informed approach to the problem — one that focused on how machines could genuinely understand and respond to the texture of human communication rather than simply pattern-matching against a predetermined FAQ. The F&B vertical was the chosen starting point: high conversation volume, clear transactional intent, and a customer base (restaurants) underserved by enterprise software that was designed for much larger operations.

VOXeat's traction validated the technical thesis. Reaching over 700 restaurant clients across Malaysia and Singapore — including recognisable brands like Baskin-Robbins, A&W, and OldTown White Coffee — demonstrated both that the product worked and that the conversational AI approach could be sold to businesses not accustomed to deploying AI in their customer-facing operations.

The pivot to Sharly: the same capability, a far larger problem

The emergence of large language models in 2022 and 2023 created both a threat and an opportunity for VOX. The threat was commoditisation: the conversational interface layer that VOXeat had built was becoming dramatically easier to replicate as generic LLM APIs became available to any developer. The opportunity was that the team's deep understanding of how to apply language models reliably — grounded in sources, resistant to hallucination, tuned for specific contexts — was precisely the capability that the new wave of document AI applications required.

The specific problem Sharly addresses is one that every knowledge worker encounters daily: the gap between the information that exists in an organisation's documents and the ability of its people to efficiently access, verify, and act on that information. Policy manuals, research papers, legal contracts, clinical trial reports, investor documents — these exist in formats that require significant human reading time to process, and that time has a real cost. Sharly's approach, built on Retrieval-Augmented Generation (RAG), lets users upload those documents and interrogate them conversationally: ask a question, get an answer, and — critically — see exactly which page of which document that answer came from. The citation is not an afterthought. It is the product's primary trust mechanism.

The differentiation from generic "chat with your PDF" tools is meaningful. Sharly is built for multi-document analysis — comparing claims across sources, identifying conflicts between documents, extracting structured data across a document set simultaneously — and for collaborative team use, with shared workspaces, annotation tools, role-based permissions, and audit logs. That combination of source rigour and collaborative infrastructure targets the professional and enterprise user rather than the casual consumer, which is both a stronger monetisation position and a more defensible product category.

Why source-grounded AI is the right architecture for professional use

The hallucination problem in large language models is not primarily a consumer inconvenience — it is a professional liability. A compliance officer who acts on a contract clause that the AI fabricated, a researcher who cites a finding the model invented, an analyst who builds a model on a data point that never existed in the source document: these are not hypothetical failure modes. They are the reason that AI adoption in regulated industries has been slower than the technology's capabilities would otherwise predict.

Sharly's docs-only reasoning mode is a direct architectural response to this concern: when enabled, the system only draws on the documents the user has uploaded, and every response links to the specific page that supports it. This is not a feature — it is a trust infrastructure that makes the tool appropriate for contexts where a wrong answer has real consequences. APA, MLA, and Chicago citation format support, inline page references, and conflict detection across documents are all implementations of the same design principle: every claim should be verifiable by the user without leaving the platform.

That architecture also defines the product's enterprise value proposition. AES-256 encryption, TLS 1.3 in transit, role-based access control, activity logs, and SSO/SAML support are table stakes for enterprise procurement. The fact that Sharly has invested in this infrastructure at its current stage — when it could have grown faster by deferring security investment — signals an understanding of where the product needs to go to serve its highest-value customer segments: universities, think tanks, compliance teams, and professional services organisations handling sensitive proprietary information.

The team: technical credibility through the pivot

Simone Macario's profile as a founder is unusually international for the Southeast Asian startup ecosystem — computer engineering background, career experience across Italy, China, and Singapore before building VOX in Malaysia. That international exposure matters for a product now competing in a global market against well-funded alternatives. Understanding how knowledge workers in different institutional contexts approach document-heavy workflows is not trivially replicable from a purely local vantage point.

CTO Davide Selvaggio's framing of the technical challenge is instructive: the value is not in connecting to LLM APIs — those are commodities — but in fine-tuning the system to handle thousands of pages of source material while staying grounded in factual knowledge rather than drifting into model-generated confabulation. That is the hard problem, and it is the problem the team spent years working on in the VOXeat context before applying it to document intelligence. The intellectual capital transfers even when the application doesn't.

The Malaysia Digital (MD) Status recognition — designating VOX as one of the leading companies driving digital transformation across Malaysia's public and private sectors — and the participation in MDEC's investor gatherings provide evidence of the government ecosystem relationships the team has built, which can be a meaningful advantage in enterprise sales to public sector institutions.

What would have to be true for this not to work

The "chat with your documents" category has attracted significant competition, including well-capitalised players like Notion AI, Microsoft Copilot, and several venture-backed startups. The category risk is that document AI becomes a feature within existing productivity platforms rather than a standalone product — which would compress the addressable market for dedicated tools like Sharly to the enterprise tier, where switching costs are higher and the security and compliance requirements favour purpose-built solutions over bundled features.

The pivot from F&B conversational AI to professional document research is also genuinely ambitious in its scope change. VOXeat's customers were SME restaurant operators in Malaysia and Singapore; Sharly's target customer is a global knowledge worker — researcher, compliance officer, analyst, academic — often within an institutional context. The sales motion, the onboarding, the competitive set, and the customer success requirements are all substantially different. Building the go-to-market capability for the new product while maintaining the technical velocity to stay ahead of the competition is the fundamental execution challenge.

We remain invested because the 2M+ user figure suggests genuine product-market fit at the individual user level, and because the enterprise feature investment signals a credible path toward the higher-value institutional tier. The team has proven it can navigate technical complexity and product transitions. In a category where the technical bar is rising fast, those are the capabilities that matter most.