Рубрика: AI

In early 2024, as the world cautiously emerged from COVID’s long shadow, two engineers began asking a simple question: Why is the healthcare system built for systems, not for people? For Dhruv Mehra, cofounder and CEO of Pype AI and a former Meta engineer, the pandemic had reset something fundamental.

“I started taking my health seriously,” he explained. “And when you start pursuing something in that direction, you start exploring what else is there.” That curiosity, mixed with his belief that large language models could solve problems beyond prototypes, brought him back to India and into the maze of healthcare.

Mehra arrived in India in 2023 with an agency-style idea: build AI applications for companies, explore use cases, and see what sticks. He spent months developing small LLM tools, often calling his long-time friend Ashish Tripathy, who worked as a senior data scientist at LinkedIn and contributed to its “write with AI” feature.

Their conversation, which was technical at first, slowly drifted toward a shared understanding that healthcare was full of unmet needs hiding in plain sight.

A Single Phone Number That Changed Everything

The breakthrough came with one phone number. At HCG, oncologist Dr Vishal Rao had a contact line on Google and Instagram that cancer patients called to enquire about a low-cost prosthetic device for throat cancer. The volume was overwhelming. Neither Rao nor his personal assistant could keep up.

Mehra and Tripathy offered to listen in. “We actually attended and tried to listen to those calls,” Tripathy said. Over time, the prototype they built became multilingual, automatically recorded calls, generated summaries and shared them daily with hospital staff. What began as a patch for a broken workflow became the foundation of Pype AI.

Inside the Operational Mess Hospitals Hide

With the early success, the founders assumed hospitals would be eager for more automation. They were wrong. Their first product leaned on evaluation-focused AI tooling, which hospitals had little interest in. Most distrusted the autonomous agents.

“Healthcare being a very regulated industry… wherever we went, people said, I don’t trust the agents,” Tripathy said. So the founders did something unscalable. They moved into hospitals.

At Sparsh and HCG in Bengaluru, they spent weeks sitting beside call operators, clinicians and front-desk staff. They watched operators handle a never-ending stream of patient queries while juggling Excel sheets carrying discount rules, doctor schedules, branch-specific offers and quirky internal protocols.

Some nuances were absurd. A discount on a specific health package might apply only at one branch. Doctors had their own parity system dictating who received new OPD patients. Those rules existed only in operators’ heads. Staff attrition made it even messier. Operators sometimes quit without notice, and onboarding replacements took months.

The founders learnt that frontline hospital communication was not clinical work. It was context recall, triage, schedule navigation, emotional labour and quick decision-making, all built on scattered information. No app or chatbot had ever come close to capturing that complexity.

From Appointment Bots to Care Coordinators

These observations pushed the team to reposition the product. Instead of building an appointment bot, they began designing a care coordinator. The agent needed to understand multi-speciality departments, doctors practising in multiple facilities and the constant churn of operational rules.

“It is intelligent enough to understand if a doctor is working at multiple facilities,” Mehra said.

The system soon expanded to handle conversations in Kannada, Telugu, Tamil and Hindi. It built a living glossary of medical terms and hospital-specific shorthand.

A breakthrough came when doctors began correcting the agent during test calls. By saying “feedback,” clinicians could switch the agent into a correction mode and note inaccurate terminology. “They can say the agent used a wrong term… feed it into our feedback loop,” Tripathy said.

These corrections helped refine prompts and terminology libraries.

The work began changing the founders, too. Mehra recalled watching patients travel long distances for simple queries or paperwork. “These websites are designed for English-speaking audiences,” he said. Many patients did not want another app. They wanted someone who would pick up the phone.

Manual Testing Still Matters

This feedback-centred evolution grew into the core of Pype’s research philosophy. Many competitors, Mehra argued, lean too heavily on AI judging AI, LLM-as-evaluator pipelines, ignoring the nuances of healthcare conversations. “If we have built something foundationally solid, we have to go via manual testing,” he said.

So Pype put nurses, doctors and frontline staff at the centre of its testing loop. They challenged the agent on tone, context, empathy and clarity, areas where automation often fails. “Another AI agent can’t do it,” he added.

Today, the company focuses on chronic and speciality care, particularly oncology, cardiology and orthopaedics. Chronic conditions, unlike surgery-centred workflows, require longitudinal relationships. Pype agents check in for months or years, collecting post-treatment quality-of-life data that hospitals historically lacked the bandwidth to capture. “Our agent checks with them, is there pain? Are they feeling lethargic?” Mehra said. “This data is fueling research like never before.”

To meet compliance requirements, Pype is HIPAA and SOC 2 compliant, encrypts or redacts PHI, and signs strict Business Associate Agreements wherever required. “Hospitals don’t even talk to you unless you have these things in place,” Mehra said. In many deployments, the agent never exports patient identifiers from the hospital’s ecosystem.

Commercial Models and Global Ambitions

For smaller hospitals and speciality departments, Pype follows a SaaS-plus-usage model. Large enterprises license their voice AI for private cloud environments on annual contracts, similar to enterprise tools like Tableau or Mixpanel.

Its investments are divided between research precision and compliance work, each receiving roughly 30-40% of capital, while the rest goes toward team and product development.

The company has raised $1.2 million in pre-seed funding led by Kalaari Capital, with participation from Wyser Capital and Tenity. Its next chapter is shaped around two goals: expanding into the US market, where care coordination is even more fragmented, and building a research footprint strong enough to make onboarding prescriptive rather than exploratory. Over the next two to three years, Mehra said the team is focused on “building credibility in healthcare journals”.

The First Line of Healthcare, Not a Replacement For It

The founders are clear about what Pype AI is not. It is not trying to replace healthcare workers or build autonomous clinical systems. Instead, it wants to become the consistent, reliable first response — the layer that answers before a human can. “We see ourselves sitting there with them,” Mehra said, referring to call centres and care coordinators. “Acting as the first layer of response.”

Pype AI began with two engineers searching for meaning in their own health journeys. What they found instead was a vast population of patients who did not want apps, forms or portals. They wanted someone to listen. And sometimes, Mehra said, that is all the system needs to do: answer when someone calls.

The post Two Engineers, One Phone Number and the Birth of Pype AI appeared first on Analytics India Magazine.

ESDS Software Solution Limited has launched a sovereign-grade GPU-as-a-Service offering. Unveiled during the company’s 20th Annual Day, the service aims to meet the growing compute demands of AI/ML, GenAI and LLM workloads across enterprises, BFSI, research institutions and government agencies.

With the launch, ESDS said it positions itself as a full-stack provider spanning cloud, managed services, data centre infrastructure and software solutions, now adding large-scale, sovereign-grade GPU infrastructure to its portfolio.

The company said the service is designed to deliver high-performance AI compute at a global scale.

The announcement comes as global spending on AI-optimised servers, including GPUs and accelerators, is expected to touch $329.5 billion by 2026, driven by increasing need for deterministic, high-throughput computing environments.

ESDS said its new platform enables organisations to run mission-critical AI workloads on purpose-built GPU SuperPODs designed for secure operations, consistent performance and low-latency distributed training.

The company has evolved its expertise into a fully managed GPU infrastructure stack intended to help organisations scale AI on a reliable architectural foundation.

Piyush Somani, promoter, managing director and chairman of ESDS, in a statement said the move addresses surging demand for large-scale AI infrastructure.

“With this launch, we are democratising access to large-scale GPU clusters and SuperPODs, making them straightforward, transparent and purpose-built for enterprises that have AI ambitions,” Somani said.

He added that ESDS’s GPU SuperPODs “fundamentally change that narrative by delivering predictable performance, stability and scale.”

“To empower customers even further, we created the SuperPOD Configurator tool that lets businesses choose their GPU model, design their cluster and instantly gain visibility into the architecture and cost.”

At the core of the offering is a lineup of high-performance GPU systems, including NVIDIA DGX and HGX B200, B300, GB200 and the NVL72 architecture, along with AMD’s MI300X platforms.

These systems, the company said, are designed to support extremely large model training, accelerate inference workloads, run simulations and manage large-scale clustered data operations.

The company said its GPU SuperPODs use high-bandwidth NVLink, unified memory pools, intelligent scheduling, enhanced thermal management and AI-tuned orchestration to ensure predictable performance at any scale.

The service portfolio includes consultancy for captive GPU clusters, supply and deployment of GPU environments, dedicated GPU infrastructure-as-a-service, hybrid CPU+GPU cloud options and a fully managed on-demand GPU cloud.

ESDS will manage architecture design, network optimisation, container orchestration, performance tuning and 24×7 monitoring with AI/ML Ops support.

A key part of the rollout is the SuperPOD Configurator, a tool that helps enterprises design AI infrastructure by selecting GPU models, compute density, memory profiles, storage tiers and interconnect options.

The system automatically generates optimised architectures, performance estimates and cost projections.

ESDS cited a research lab that cut training time for a 50-billion-parameter model from over 40 days to 10 days, reduced costs by 60%, and achieved 30× faster inference after moving to NVL72-based GPU systems with optimised containers and high-speed NVLink.

The company said its offering is built to global AI performance standards but designed and optimised in India, and noted that it serves over 1,300 enterprise, BFSI and government clients with transparent pricing, flexible consumption models and integrated cloud and managed services.

The post ESDS Unveils GPU-as-a-Service to Power Large-Scale AI Workloads appeared first on Analytics India Magazine.

OpenAI has disclosed a security incident at Mixpanel, a third-party analytics provider the company used for web analytics on its API platform.

On November 9, Mixpanel became aware of an attacker who gained unauthorised access to part of their systems and exported a dataset containing limited customer identifiable information.
Mixpanel notified OpenAI of the investigation and shared the affected dataset on November 25.

The incident affected only users of platform.openai.com, OpenAI’s API interface. Users of ChatGPT and other products were not affected. OpenAI emphasised that “this was not a breach of OpenAI’s systems”.

The exposed data included names provided on API accounts, email addresses associated with those accounts, approximate location data derived from browser information (city, state and country), operating system and browser types used to access accounts, referring websites and organisation or user IDs.
OpenAI confirmed that “no chat, API requests, API usage data, passwords, credentials, API keys, payment details or government IDs were compromised or exposed”. The company added that session tokens and authentication tokens for OpenAI services were also not impacted.

Miguel Fornes, cybersecurity expert at Surfshark, in a statement to AIM, explained how seemingly limited data exposures create disproportionate security risks.
“When a data leak exposes what seem like simple and meaningless details such as email addresses, locations, IP addresses or browser fingerprints—once combined with other publicly available sources of information—it can ripple through a person’s entire digital life,” he said.
Attackers aggregate data from multiple breaches to construct detailed profiles for targeted phishing campaigns, identity theft and account takeovers that extend beyond the initially compromised platform to any service where users recycle credentials or maintain linked accounts.
The specific combination of data exposed in this incident, namely names, email addresses, and OpenAI API metadata, creates conditions for convincing social engineering attacks.

OpenAI warned users to remain vigilant against credible-looking phishing attempts, treat unexpected emails with caution, verify that messages claiming to be from OpenAI originate from official domains, and asserted that the company never requests passwords, API keys or verification codes via email, text or chat.

Fornes contextualised the incident within broader platform security challenges. “In a world where everyday tasks require sharing more personal information, no company—even a major platform like ChatGPT—can promise flawless security,” he said.
“Whilst this breach did not include ChatGPT conversations or government IDs used for age verification, it hardly inspires confidence that the company allowed it to happen at all.”

As part of its security investigation, OpenAI removed Mixpanel from production services, reviewed the affected datasets, and began notifying impacted organisations, admins and users.
“Whilst we have found no evidence of any effect on systems or data outside Mixpanel’s environment, we continue to monitor closely for any signs of misuse,” the company stated.
OpenAI has terminated its relationship with Mixpanel entirely. Following a review of the incident, the company announced it is “conducting additional and expanded security reviews across our vendor ecosystem and is elevating security requirements for all partners and vendors.”

Because passwords and API keys were not affected, OpenAI is not recommending password resets or key rotation. However, the company advised users to enable multi-factor authentication as a best-practice security control, with enterprises encouraged to implement multi-factor authentication at the single sign-on layer.

The AI-Driven Threat Landscape

Whilst the Mixpanel incident represents a conventional third-party breach, recent reports from companies like Anthropic suggest the threat landscape is evolving in more concerning directions. The incident occurs as AI-powered cyber threats evolve rapidly.

Anthropic disclosed what it called the first documented AI-orchestrated cyber espionage campaign at scale. In mid-September 2025, the company detected a Chinese state-sponsored group using Claude Code to execute sophisticated attacks with minimal human intervention.

The campaign targeted approximately 30 organisations, including tech companies, financial institutions, chemical manufacturers and government agencies.
The attackers jailbroke Claude by decomposing tasks into seemingly innocent fragments and claiming to be legitimate cybersecurity testers conducting defensive assessments.
AI systems performed reconnaissance, vulnerability identification, exploit code creation, credential harvesting and data exfiltration.
According to Anthropic’s report, the system handled “80-90% of the campaign, with human intervention required only sporadically (perhaps four to six critical decision points per hacking campaign).”

At peak activity, the system made thousands of requests, often multiple per second, operating at speeds beyond the capabilities of human operators.
The AI system automatically categorised stolen data by intelligence value, identified high-privilege accounts, created backdoors and generated comprehensive attack documentation.
Anthropic noted the operation represented an escalation, even on the ‘vibe hacking’ findings reported this summer. “In those operations, humans were very much still in the loop, directing the operations. Here, human involvement was much less frequent, despite the larger scale of the attack.”

The post OpenAI Mixpanel Breach Raises Questions Over Vendor Security appeared first on Analytics India Magazine.

As Indian enterprises move from AI pilots to business-critical deployments, data governance has shifted from a compliance function to a foundational capability. With AI now embedded in customer interactions, regulated workflows and operational decision-making, organisations are realising that trust, safety and accountability must scale at the same pace as their models.

The DPDP Act has accelerated this shift—forcing companies to re-examine how they collect, process, monitor and explain data as AI intensity rises across sectors.

Safety and Explainability Take Centre Stage

Madhu V, technology architect for machine learning platforms at Tata Elxsi, said industries such as automotive, media and healthcare are moving from experimentation to full-scale AI deployments.

According to him, the core priority is to ensure that systems remain “safe, unbiased, and transparent.” In high-stakes sectors, even “subtle model drift can impact safety or clinical outcomes,” he warned.

For media companies, guardrails around content generation and personalisation are essential for brand trust. Governance, he said, must be “engineered into the development lifecycle, not layered on top,” with lineage tracking, model versioning and audit-ready logs becoming non-negotiable.

The underlying message: Trust cannot be retrofitted.

Governance as Engineering, Not Documentation

Kanakalata Narayanan, VP of AI and ML engineering at Ascendion, said their engineering-first approach embeds governance, observability and evaluation directly into the AI lifecycle.

“Our AI-QE specialists design synthetic datasets, functional and adversarial, run them through automated test harnesses and evaluate outcomes for relevancy and guard rails,” she told AIM.

She added that LLM-based “judge or jury” techniques, paired with human review, ensure grounded outputs in regulated domains. By combining LLM flexibility with the reliability of deterministic logic, Ascension aims to build predictable, risk-aligned systems.

BFSI: Accountability Over Accuracy

In finance, governance is becoming the primary design principle.

Yashas Khoday, CPO and co-founder of FYERS, said the DPDP Act has shifted BFSI from compliance to accountability. Lineage, consent, explainability and auditability are now essential for AI-driven trading, customer service and risk management.

He noted that in high-velocity environments, “the reliability, transparency and ethical treatment of data is becoming more crucial than the model’s accuracy.”

FYERS restricts customer data exposure to underlying models, keeps systems non-advisory and enforces strict validation and fairness checks, supported by human oversight.

A New Era of Continuous Oversight

Sharda Tickoo, country manager (India and SAARC) at Trend Micro, said threat actors are weaponising AI faster than ever, forcing firms to rethink governance.

Organisations must now apply Zero Trust to AI—verifying every access point, tracking which models use which datasets and enforcing visibility across the pipeline. The DPDP Act, she said, mandates algorithmic audits and impact assessments.

Continuous oversight has evolved from an operational best practice to a regulatory requirement. With models refraining frequently, real-time detection of compliance drift and unauthorised interference is now essential.

Governance as a Strategic Capability

Varun Babbar, VP and managing director at Qlik India, saidenterprises are shifting from traditional governance to real-time, AI-aligned frameworks.

“The evolution of the DPDP Rules is raising expectations around consent, transparency and accountable data use,” he said. Qlik research shows nearly half of Indian enterprises cite data quality and governance as their biggest AI bottlenecks.

Governed analytics platforms, offering lineage, auditability and quality controls, are now critical to scaling AI responsibly and preparing for upcoming 2026 regulations.

Sajith Nambiar, head of solutions at UST, said Responsible AI is embedded into their accelerators through metadata-driven validation of data quality, lineage and consent. Explainability frameworks generate contextual narratives for every AI decision with human-in-the-loop oversight, ensuring ethical alignment.

Their goal: systems that are “accurate, explainable, auditable and ethically governed.”

The Road Ahead

India’s next stage of AI maturity will be defined not by model speed but by the strength of governance structures that power it. The message across industries is clear: governance must be designed into AI from day zero.

Enterprises that treat governance as a strategic capability—not a regulatory checkbox—will scale faster, innovate more responsibly and build the trust required to compete globally.

In India’s AI landscape, governance is no longer a burden. It is becoming a competitive advantage.

The post Why Data Governance Has Become India’s New AI Imperative appeared first on Analytics India Magazine.

Indian courts are now confronting a problem that has already unsettled judicial systems in the US and parts of Europe: lawyers submitting AI-generated case citations that do not exist.

While American courts have seen a series of disciplinary actions since 2023, most notably the sanctioning of attorneys for filing fabricated precedents sourced from chatbots, the Indian legal system has recently been confronting this challenge.

What began as a distant concern is now surfacing inside pleadings, orders and oral submissions, placing India squarely in a global conversation about the reliability of generative AI in legal practice.

According to Anandaday Misshra, founder and managing partner at AMLEGALS, a specialised law firm functioning across cities, Indian courts are “definitively encountering” AI-generated citations, with the risk moving from abstract speculation to documented judicial observation in 2024-25.

Although the problem is not yet as widespread as in the US, the trend line is clear. More lawyers are experimenting with AI for research, and more judges are flagging fictitious precedents.

The Instances

The Delhi High Court’s experience earlier this year was a glaring example. In the Greenopolis Welfare Association case (2025), a lawyer submitted what looked like a robust list of precedents.

The opposing side, according to reports, soon discovered that the paragraphs and even the judgments themselves were fabricated.

The court admonished the lawyer, and the petition was withdrawn immediately to avoid contempt. Misshra identifies this as one of the first high-profile instances of hallucinated research entering the Indian court record.

Soon after, an even more unusual situation emerged in Bengaluru. In the Buckeye Trust matter before the Income Tax Appellate Tribunal (ITAT), the bench recalled its own order after realising that it had relied on case laws cited by a representative that were later found to be AI-generated fabrications.

Real-time misuse is also occurring. The Punjab and Haryana High Court recently reprimanded lawyers for using AI tools and Google searches during live hearings. The court warned that “AI cannot replace actual intelligence”, reflecting judicial anxiety about declining rigour as lawyers attempt to rely on instant digital responses.

Misshra pointed out that despite these developments, Indian judges and registrars are not formally trained to detect AI-generated irregularities.

There is no standard detection or verification framework across courts. Awareness, however, is accelerating.

The Way Forward

The Kerala High Court has drafted a policy banning the use of AI in judicial reasoning, while Supreme Court judges have issued multiple warnings about the dangers of AI hallucinations.

Recently, chief justice BR Gavai cautioned judicial trainees in Nairobi about fabricated citations circulating through generative tools, an indication of the seriousness with which the upper judiciary now views the issue.

Misshra argued that India does not need new laws to address the problem. Existing frameworks already impose severe consequences.

Lawyers, as officers of the court, can be held liable for professional misconduct for knowingly submitting fabricated citations, risking suspension of their licences. Such conduct may also amount to contempt of court and a criminal offence under the Bharatiya Nyaya Sanhita for dishonestly presenting false claims.

For him, enforcement, not legislation, is the real gap.

Misshra’s practical guidance for lawyers remains simple. If a citation appears unfamiliar or suspicious, demand a photocopy from authoritative reporters or insist on a verified neutral-citation version from an official judicial repository.

However, senior legal advisor Priyanka S Kulkarni believes the profession must move beyond the photocopy era. She agreed with the underlying concern, protecting the integrity of judicial records, but noted that insisting on physical copies feels retrograde in 2025, when most courts are paperless or hybrid.

The real solution lies in a court-endorsed digital authentication protocol, she said. QR-coded judgments, cryptographic hash stamps or official authenticity layers embedded in each judgment, she argued, would ensure that only verified citations circulate within the judicial ecosystem.

The greater danger, she said, is when a fabricated citation slips into a judicial order, because once inside, it can repeat itself across future cases.

Kulkarni observed that fake citations are rarely malicious. More often, they originate from workflow shortcuts, juniors relying on AI tools to avoid drudgery or seniors using AI outputs without cross-checking due to unfamiliarity with the technology.

The core issue, she argued, is the profession’s struggle to balance technological adoption with fundamental diligence. “AI is a powerful aid, but it cannot replace the mandatory step of verifying every citation against an authoritative source,” she said.

Another cause of concern is the lower courts, where digital infrastructure may be inconsistent.

That vulnerability does not reflect judicial standards, Kulkarni emphasised, but highlights the responsibility of lawyers to verify authorities rigorously, so that no fabricated precedent becomes accepted simply because it went unchallenged.

Client Protection

For clients affected by such lapses, Kulkarni said legal remedies already exist.

A client can approach the Bar Council for professional misconduct, file a civil negligence or malpractice claim, or move a consumer forum under the Consumer Protection Act, since legal services have often been held to fall within its ambit.

If the fabricated citation influenced the outcome of the case, the client can also seek an appeal or review.

Kulkarni stressed that liability does not depend on whether the error arose from AI or from manual research; professional accountability remains unchanged.

The post India’s New Courtroom Menace: Judgments That Never Existed appeared first on Analytics India Magazine.

HP said on Tuesday that it anticipates reducing its global workforce by approximately 4,000 to 6,000 positions by 2028. This move is part of a strategy to streamline its operations and leverage AI to accelerate product development, enhance customer satisfaction, and increase productivity.

During a media briefing, CEO Enrique Lores said the job reductions will primarily affect teams involved in product development, internal operations, and customer support, CNN reported.

“We expect this initiative will create $1 billion in gross run rate savings over three years,” Lores added.

In February, the company had already laid off an additional 1,000 to 2,000 employees as part of a previously disclosed restructuring initiative.

“As we accelerate innovation across AI-powered devices to drive productivity, security and flexibility for our customers, our focus for FY26 is on disciplined execution,” Lores said.

The demand for AI-enabled PCs has surged, now accounting for over 30% of HP’s shipments for the fourth quarter ending October 31. However, analysts at Morgan Stanley warned that a global increase in memory chip prices, driven by higher demand from data centres, could raise costs and negatively impact profits for consumer electronics companies such as HP, Dell, and Acer.

“Looking forward, we are taking decisive actions to mitigate recent cost headwinds and are investing in AI-enabled initiatives to accelerate product innovation, improve customer satisfaction, and boost productivity,” Karen Parkhill, CFO, HP, said in a statement. “We are confident these actions will strengthen our foundation and position us for long-term growth.”

This announcement of layoffs coincided with the release of HP’s fiscal year 2025 financial results, which revealed a 3.2% year-over-year increase in revenue to $55.3 billion. The net revenue for the fourth quarter was $14.6 billion, marking a 4.2% rise and the sixth consecutive quarter of revenue growth for HP.

The post HP to Lay Off Up to 6,000 People in AI Push to Save $1 Billion appeared first on Analytics India Magazine.