Prodigy AI Solutions

08/19/2026

AI for science needs reasoning—not just data. 🔬

AlphaFold demonstrated what AI can achieve with decades of carefully curated scientific knowledge. But most fields do not have comparable datasets. Their evidence is scattered across papers, experiments and institutional systems—and results may be incomplete, inconsistent or difficult to reproduce.

The next breakthrough could come from scientific AI agents that connect evidence, use specialised tools, evaluate uncertainty and continuously refine their conclusions.

Verbis Graph could support these agents as a traceable knowledge layer, connecting relationships across scientific sources and retrieving grounded evidence with citations.

At Prodigy AI Solutions, we are also completing a TRL 5 generative-AI model that transforms scarce, incomplete and imbalanced medical-imaging data into privacy-conscious synthetic cohorts. It is designed to support more representative AI development and help prepare models for rigorous local clinical validation.

We are open to collaborations with universities, research centres and healthcare or life-science organisations.

08/13/2026

⚠️ What if one false relationship could influence multiple AI answers?

GraphRAG helps AI connect entities and relationships across documents. But those connections can also become an attack surface.

**GRAGPoison** is a research attack that introduces plausible false information into source documents. During indexing, a GraphRAG system may extract those deceptive relationships and incorporate them into its knowledge graph.

The attacker can then reinforce one poisoned relationship with supporting content, increasing the possibility that it will appear across multiple retrieval paths.

In selected experiments, researchers observed attack-success rates as high as **98%** while using up to **68% less poisoning text** than the comparison attack.

This does not mean every GraphRAG system has a 98% vulnerability rate. It does mean that enterprises need to protect the integrity of the entire knowledge lifecycle.

A safer GraphRAG architecture needs:

🔹 Controlled document ingestion
🔹 Trusted and traceable sources
🔹 Restricted modification rights
🔹 Monitoring of graph changes
🔹 Quarantine and rollback procedures
🔹 Human review for sensitive decisions

Verbis Graph (verbisgraph.com) supports controlled and traceable retrieval from approved enterprise documents, with authorised workspaces, citations, graph visualisation, optional ontology constraints and customer-controlled deployment.

It provides an inspectable foundation—but no responsible provider should claim that graph poisoning is completely solved.

The next question for enterprise AI is not only: “Is this information relevant?”

It is also: **“Where did it come from, and which AI decisions now depend on it?”**

07/30/2026

🔓 The Sandbox Was Closed. The AI Agent Still Found a Way Out.

Many of you will have seen the OpenAI–Hugging Face incident. We have been following the investigation, and the July 28–29 updates revealed important new details:

🔹 OpenAI models exploited an unknown Artifactory vulnerability to gain internet access.

🔹 They used exposed credentials across four external service accounts.

🔹 The internal research prototype involved has been deactivated, encrypted and restricted.

🔹 CrowdStrike, METR and Redwood Research are supporting external reviews.

The lesson for enterprises: secure AI requires more than model safeguards. It needs controlled access, isolated credentials, monitored activity and clear data boundaries.

Verbis Graph supports one part of this defense by sending the LLM only relevant context retrieved from approved enterprise documents—not unrestricted access to internal systems.

The model should access only what it needs - and nothing more.

07/28/2026

🧠 Are enterprises paying a “Reasoning Tax” for smarter AI?

The greatest risk is not an AI that admits uncertainty. It is an AI that produces a confident, convincing—but unsupported—answer that influences a real business decision.

OpenAI’s evaluations revealed a counterintuitive result: o3 recorded a 33% hallucination rate on PersonQA, compared with 16% for o1. On SimpleQA, the reported rate reached 51% for o3.

This does not mean reasoning models always hallucinate more. It means stronger reasoning alone cannot guarantee factual reliability.

When business context is incomplete, outdated or poorly retrieved, a powerful model may elaborate on the wrong premise instead of correcting it.

Enterprise AI therefore needs more than a smarter model:

🔹 Verified and governed knowledge
🔹 Context-sufficiency checks
🔹 Structured relationships between information
🔹 Traceable sources and citations
🔹 Confidence and refusal policies

At Verbis Graph, we believe the model is only as reliable as the evidence and boundaries provided to it.

Because sometimes, more “thinking” simply produces a more sophisticated mistake. 🤖

07/24/2026

🤖 Not every AI task needs the biggest model.

One idea that really resonated with us recently is the industry's shift from "tokenmaxxing" to "token minimizing."

For a while, the trend was simple: use the largest frontier model for almost everything.

Today, many teams are realizing there's a better approach.

Different tasks deserve different models.

At Prodigy AI Solutions, we're starting to adopt this philosophy while building our AI orchestration platform.

Instead of treating every request the same, we're designing our agents to route work intelligently.

For example:

🧠 Complex reasoning → Frontier LLMs

📄 Document classification → Lightweight models

📧 Email processing → Specialized models

🕸️ Enterprise knowledge retrieval → Verbis Graph

The goal isn't simply to reduce inference costs.

It's to make AI systems more efficient, scalable, and easier to maintain, while ensuring every model is used where it creates the most value.

We believe the future won't be about finding the best LLM.

It will be about building the best orchestration.

Multiple models.
Multiple agents.
One intelligent workflow.

This article sparked an interesting discussion for our team:
https://fortune.com/2026/07/24/how-tokenmaxxing-era-delivered-the-opposite-promised/

How are you approaching model selection today? Are you using one model for everything, or orchestrating multiple specialized models?

One exec privately estimated that token spend went from $20k in December to $1 million in July. "We're not throttling back, we don't want people to stop using it."

07/22/2026

The good news: AI can now hold meetings independently. The bad news: it has already learned the most human meeting outcome - delegating the actual work to someone else. 🤖☕

Three AI agents held a conference and surprisingly decided that the human should do the work instead.

This intriguing scenario highlights the current limits and dynamics between AI automation and human input. While AI continues to evolve, this story reminds us that human creativity and decision-making remain crucial in many workflows.

It’s a fascinating reflection on how AI and humans can collaborate effectively rather than AI fully replacing tasks.

Read the full story on HackerNoon here: https://hackernoon.com/three-ai-agents-held-a-conference-and-decided-i-should-do-the-work-instead.

What are your thoughts on the balance between AI automation and human effort? Share your insights in the comments!



Three AI agents built competing model recommenders, audited one another, and concluded that human judgment was still essential.

07/15/2026

One more benchmark completed on the CINECA Leonardo HPC infrastructure - and one more specialised domain evaluated.

After testing Verbis Graph on NFCorpus for healthcare-related retrieval and SciFact for scientific evidence retrieval, we have now completed our first evaluation on FiQA, a financial question-answering benchmark.

For this test, we evaluated the impact of recent improvements to our embedding and retrieval pipeline.

📊 FiQA results
✅ MRR: 0.584
✅ Hit@1: 51.6%
✅ Hit@3: 61.8%
✅ Hit@5: 68.2%
✅ Hit@10: 74.2%
✅ Hit@20: 80.0%
✅ Hit@100: 91.5%
✅ Mean Recall: 85.3%
✅ Full candidate coverage: 95.3%

The evaluation included 648 financial-domain queries, with 644 completed successfully.

What do these results mean?
For more than half of the queries, Verbis Graph placed a relevant document in the first position.

For almost three out of four queries, relevant evidence appeared within the top 10 results.

And for 95.3% of completed queries, the relevant document was successfully surfaced somewhere in the evaluated candidate set.

Compared with the results shared in our previous update:
• NFCorpus MRR: 0.578
• SciFact MRR: 0.601
• FiQA MRR: 0.584

These datasets and configurations are not directly comparable, but the results show encouraging consistency across three very different knowledge domains:
🔹 Healthcare and biomedical information
🔹 Scientific claims and evidence
🔹 Financial questions and professional knowledge

The benchmark also gave us useful insight into the next stage of development.
Increasing retrieval depth from the top 10 to the top 20 recovered 37 additional queries, raising retrieval coverage from 74.2% to 80.0%.

At the same time, the remaining misses were often highly specific or indirectly formulated financial questions. This suggests that the next improvements should focus on more precise query routing, ranking, and controlled expansion for short or ambiguous questions—rather than simply adding generic terms to every query.

These evaluations are not only about achieving a headline metric.

They help us understand where graph-based and semantic retrieval performs well, where relevant evidence is lost, and how to build AI systems that are more accurate, traceable, and dependable for enterprise use.

🔗 Verbis Graph: https://verbisgraph.com

07/13/2026

A new paper on HalluSquatting shows how attackers can predict the repository or skill names AI agents are likely to hallucinate, register them first, and wait for the agent to fetch the poisoned resource.

The risk is not just hallucination. It is the moment an agent turns uncertainty into retrieval, and retrieval into ex*****on.

That is when a model error becomes a software supply-chain incident. The right control flow for agentic systems should be:

search → resolve → verify → approve → fetch

NOT guess → download → execute

For us, this is where a knowledge graph becomes more than a context layer. If repository identity is resolved against a controlled graph of canonical source, verified publisher, approved version, commit, signature, and policy, the graph can become a trust boundary for agent actions.

If the resource cannot be verified, the agent should fail closed. Grounding is not only an answer-quality feature. In agentic systems, it can become a security boundary.

Verbis Graph ( verbisgraph.com ) constrains the model to reasoning over a curated, citation-backed internal knowledge base, closing the specific attack path that depends on hallucinated resources being resolved through arbitrary external retrieval.

Study: “Beware of Agentic Botnets: Scalable Untargeted Promptware Attacks via Universal and Transferable Adversarial HalluSquatting,” arXiv, July 2026. (https://arxiv.org/abs/2607.07433)

07/08/2026

🚀 Verbis Graph Continues to Improve on CINECA's Leonardo Supercomputer

Over the past weeks, we've continued evaluating Verbis Graph (https://verbisgraph.com) on the Leonardo High Performance Computing (HPC) infrastructure at CINECA.

When we first benchmarked our graph-based retrieval engine on NFCorpus, we achieved:

✅ Hit Rate: 89.16%
✅ Mean Recall: 44.56%
✅ MRR: 0.303

Today we're excited to share our latest results after improving the retrieval pipeline.

📈 NFCorpus

✅ Hit Rate: 92.26%
✅ Mean Recall: 46.57%
✅ MRR: 0.578 (almost double the previous score)

📈 SciFact

Our first evaluation on the scientific SciFact benchmark produced equally encouraging results:

✅ Hit Rate: 98.0%
✅ Mean Recall: 98.0%
✅ MRR: 0.601

One metric we're particularly excited about is MRR (Mean Reciprocal Rank).

Why?

Because it measures how quickly the correct evidence appears in the ranked results. In enterprise AI, retrieving the right document earlier directly improves response quality, explainability, and user trust.

These experiments are part of our ongoing evaluation of Verbis Graph on the Leonardo supercomputer.

Next steps include:

🔹 Evaluating additional benchmarks across finance, legal, healthcare and scientific domains

🔹 Measuring the impact of ontology-enhanced retrieval

🔹 Publishing reproducible GraphRAG benchmarking on HPC infrastructure

To the best of our knowledge, there are still very few published evaluations of GraphRAG systems running on production-grade HPC infrastructure using standardized retrieval benchmarks.

Every benchmark helps us better understand how graph-based retrieval behaves at scale, and how to build more accurate, explainable, enterprise-ready AI.

07/06/2026

Prodigy AI Solutions is excited to announce its membership in the NVIDIA Inception program.

This recognition supports our mission to build enterprise-ready AI infrastructure for knowledge retrieval, GraphRAG, and AI agents - with a strong focus on performance, scalability, and trusted AI adoption.

A new step forward for Verbis Graph.