Research & engineering
Research that ships. Engineering that holds.
Every area starts with a problem a customer hit inside their perimeter, and ends as engineering that runs there: in a product, under audit. We publish what we learn along the way.
200,000+
Open security models
downloads of our open security models
SecurityLLM is cited in the Foundation-sec-8B technical report by Cisco Foundation AI and Yale, and used as a security-specialist baseline in university research.
- Peer-reviewed
- Accepted at a NeurIPS 2026 workshop and CIKM 2026
- Open
- Models, data and code, in public since 2023
- Cited
- By Cisco Foundation AI, Yale and university research
- Shipped
- Every research area runs in a product
The journey
From one open model to a research engine.
Since December 2023, in public.
Dec 2023
First open security model
Mamba-2.8B-CyberSec, a state-space model for security, released on Hugging Face.
Jan 2024
ZySec-7B, SecurityLLM
A DPO-tuned security model across 30+ domains, since downloaded nearly 200,000 times.
Mar 2024
Built for air-gapped hardware
AWQ and GGUF builds so the model runs offline, inside the perimeter.
May 2024
Named in the literature
Included in a comprehensive review of generative AI in cybersecurity.
Oct 2024
Open safety data
The harmful_behaviors dataset, for testing refusal and misuse.
Q1 2025
Open data for retrieval
Contextual RAG datasets, crime stories, and open document models for constrained hardware.
Apr 2025
Cited by Cisco
Cisco Foundation AI and Yale University cite SecurityLLM in the Foundation-sec-8B technical report.
Jul 2025
Best Emerging AI Company
Named Best Emerging AI Company of the Year 2025 at the Indian Business Excellence Awards.
Oct 2025
A baseline for others
Used as a security-specialist model in POLAR, Binghamton University's threat-prioritisation research.
Jul 2026
RelataDB goes open source
An AI-native knowledge engine: agent memory, graph and hybrid search, with SDKs in Python, TypeScript and Go.
Sep 2026
Accepted at NeurIPS
AgentInSight, signed tool manifests and declared-versus-deployed audit, accepted as a NeurIPS 2026 workshop poster.
Nov 2026
CIKM 2026, Rome
HyperMind: claim resolution and reputation tracking for trustworthy multi-agent answers.
Now
Saqr, and what comes next
A 27.8B security model for Infinia Technologies, 200,000+ downloads across our open models, and SHABD in progress.
Research areas
Six real problems. Researched, then engineered.
Areas closest to the view you chose come first, marked “For you”.
RAG quality
Knowledge Reliability Systems
Retrieval is easy to demo and hard to trust. We study when an answer can be relied on: which source supports it, how confident the system should be, and who was right when sources or agents disagree.
The problem
Answers from retrieval looked right, but were not always traceable to a source.
What we researched
ARAI, and HyperMind's claim resolution and reputation tracking (CIKM 2026).
What we engineered
Citations to file, page and passage, and agents weighted by how often they were right.
Where it runs
AutonousAutonous IntOps
Our work
- ARAIOur retrieval-reliability engine, embedded in Autonous and Autonous IntOps.
- HyperMindClaim resolution and reputation tracking for multi-agent answers. CIKM 2026.
- Contextual RAG datasets ↗Open rewriter and relations datasets for retrieval research, plus RAG markdown documents.
Building on · 5 references
- Retrieval-Augmented Generation for Knowledge-Intensive NLP TasksLewis et al. · NeurIPS 2020
- RAGAs: Automated Evaluation of Retrieval Augmented GenerationEs et al. · EACL 2024 Demos
- Self-RAG: Learning to Retrieve, Generate, and Critique through Self-ReflectionAsai et al. · ICLR 2024
- A Survey on Truth DiscoveryLi et al. · SIGKDD Explorations 2016
- Survey of Hallucination in Natural Language GenerationJi et al. · ACM Computing Surveys 2023
Publications
Peer review is part of the product.
HyperMind: Claim Resolution and Reputation Tracking for Trustworthy Multi-Agent LLM Aggregation
Lowest Brier score on all four LLMs tested, at $0.005 per 100 questions. Every resolved claim leaves an auditable chain of custody.
AgentInSight
Enforces agent tool calls against signed manifests. In a live case study it found a server declared for one tool exposing 59.
CyberPod: Entity Recognition and Multi-hop Criminal Network Intelligence for Investigative Analysis
F1 0.87 on AttackDB corpora, a 6 to 11 point gain over off-the-shelf baselines, with 412 ms for 3-hop reasoning in production.
ZySec-7B (SecurityLLM): a DPO-tuned security model across 30+ domains ↗
Nearly 200,000 downloads, with community Spaces and quantized builds. Cited in the Foundation-sec-8B technical report by Cisco Foundation AI and Yale.
Receipts, Not Verdicts: Judge-Free Verification for Multi-Agent LLM Systems
Re-derivable receipts instead of LLM judges. Reputation weighting improved Brier score by 5.4 to 12.5% across eight LLMs.
SHABD: Source-History-Anchored Bias Detection in Media
Detecting systematic editorial framing by comparing how a source covers comparable subjects over time.
Open models and data
Built in the open, run in your perimeter.
ZySec-7B, our open security model across 30+ domains. Nearly 200,000 downloads, community Spaces and quantized builds. Apache 2.0.
ZySec-AI/SecurityLLM ModelSaqr27.8B-parameter multimodal cybersecurity model, 262K context, thinking and non-thinking modes. Reported 96.2 on CyberMetric-2000.
saqr.iamsaif.ai Open sourceRelataDBUnlimited memory for AI agents: a knowledge engine with graph, hybrid search, policy masking and drop-in protocols.
relatadb.devCited by
Other researchers build on our work.
Research partnerships
Work with our research team.
We collaborate with IIT Madras and welcome universities, labs and public-sector teams on the open problems in these six areas.