About
Articles by Saeed
Activity
8K followers
-
Saeed Kasmani, Ph.D. shared thisOpenTelemetry may end up being one of the most useful standards in the AI stack. Not because tracing is exciting. Because agent systems fail in messy, distributed ways. When an AI workflow breaks, the issue is rarely just the model. It is usually somewhere in the chain: bad retrieval, stale tool output, prompt drift, or a brittle handoff between agents. That is why traceAI caught my eye this week. It is an open-source tracing framework built on OpenTelemetry, supporting 50+ AI frameworks across Python, TypeScript, Java, and C#. The pitch is refreshingly practical: instrument your app in a few lines and send the traces to the observability stack you already use. That design choice matters. Once you have agents, tools, retrieval, memory, and multiple models in one workflow, debugging stops being an "AI problem" and starts looking a lot like distributed systems engineering. I think more of the AI stack will move this way: • Standard traces • Existing telemetry backends • Shared debugging workflows •Less vendor-shaped instrumentation That is healthier than every AI tool inventing its own black box. The teams that win in production may not be the ones with the flashiest model demos. They may be the ones that make AI systems easier to observe, compare, and fix. 🔎 🔗 https://lnkd.in/g576MZx6 #AI #OpenTelemetry #LLMObservability #MLOps #AIOpsGitHub - future-agi/traceAI: Open Source AI Tracing Framework built on Opentelemetry for AI Applications and FrameworksGitHub - future-agi/traceAI: Open Source AI Tracing Framework built on Opentelemetry for AI Applications and Frameworks
-
Saeed Kasmani, Ph.D. shared thisOpenAI has released Privacy Filter, a small redaction model with 1.5B total parameters, 50M active parameters, tunable precision/recall presets, and an Apache 2.0 license. A 128k-context privacy model is a pretty clear sign that AI data plumbing is changing. The part I like is not the launch itself. It is where this kind of model fits. Most organisations still handle sensitive text with a patchwork of regex rules, brittle masking scripts, and manual clean-up steps before anything reaches an LLM. That approach falls apart once the documents get long, messy, or inconsistent. A long-context span-labeling model gives teams another option. You can run privacy detection locally, inspect the implementation, fine-tune it for your own document patterns, and treat redaction as part of the application flow instead of a separate hygiene job. That is useful in exactly the places where AI gets hard: claims notes, case files, support logs, clinical text, and internal knowledge bases full of half-structured human writing. The deeper shift is architectural. We are moving from "strip sensitive fields if you can" to "design the system so privacy controls travel with the data path." That usually leads to better reviews, cleaner ownership between data and AI teams, and fewer ugly surprises later. This is the kind of release I pay attention to because it is small, practical, and immediately deployable. 🔐 🔗 https://lnkd.in/g99nYMbi #AI #PrivacyEngineering #LLMOps #DataArchitectureGitHub - openai/privacy-filter: OpenAI Privacy FilterGitHub - openai/privacy-filter: OpenAI Privacy Filter
-
Saeed Kasmani, Ph.D. shared thisObservability stacks are starting to care about output quality, not just uptime. 📊 Grafana's new AI Observability preview tracks inputs, outputs, execution flows, session traces, low-quality responses, policy breaches, and anomalous behaviour in the same place teams already watch their services. That is a bigger shift than it sounds. Classic telemetry tells you whether the system is alive. AI telemetry also needs to tell you whether the system is still useful after a prompt change, model swap, retrieval tweak, or tool-call chain. Those are not the same question. Once you run agents, a green dashboard can hide a broken product: • Latency looks fine • Tokens are flowing • Nothing is technically down • But answer quality is drifting and the workflow is quietly getting worse I think the best AI platforms will fuse tracing, evals, and operational monitoring into one loop. If those stay separate, teams will keep shipping agents that are technically healthy and commercially useless. That feels like the real story behind this launch. 🤖 🔗 https://lnkd.in/gZNW4qYU #AIEngineering #LLMOps #ObservabilityGrafana Labs Targets the “AI Blind Spot” with New Observability Tools Announced at GrafanaCON 2026 | Grafana LabsGrafana Labs Targets the “AI Blind Spot” with New Observability Tools Announced at GrafanaCON 2026 | Grafana Labs
-
Saeed Kasmani, Ph.D. shared thisA useful reality check from healthcare AI this week 🏥 Researchers tested 10 frontier multimodal models on 539 real inpatient cases from a South African public hospital. The surprising part was not who came first. It was how tightly clustered the models were, despite very different costs. That changes the buying and design conversation. In real deployments, the question is rarely “which model won the benchmark?” It is usually: • Can it handle the inputs you actually have? • Does it return complete outputs reliably? • What happens when context is messy? • Can you afford to run it at scale? One result jumped out at me: adding radiology reports lifted performance by 6%. So the practical lesson is not “always buy the smartest model.” It is that workflow and input design still move the needle a lot. If you work in health, banking, or government, that should sound familiar. Production systems live or die on context quality, consistency, and cost discipline, not on leaderboard bragging rights. That is the kind of AI result I trust more, because it comes from a setting where bad outputs have consequences. 🔗 https://lnkd.in/gfrZjVCy #AI #HealthcareAI #MultimodalAI #AIArchitecture #DigitalHealth
-
Saeed Kasmani, Ph.D. shared thisThree files usually describe the same product idea: a mockup, a slide, and a ticket. By the time engineering gets them, they already disagree. That is why Anthropic’s Claude Design stood out to me this week. The headline is “AI for design”. The more useful signal is the workflow underneath. Anthropic says Claude Design can read a team’s design system, generate prototypes, export to formats like PPTX and HTML, and hand the result to Claude Code as an implementation bundle. If that works well in practice, the value is not just faster mockups. It is less translation loss between product, design, and engineering. That is where a lot of delivery friction lives. Not in drawing the first screen, but in watching the original intent get diluted across handoffs, rewrites, and new artifacts. So I would not judge tools like this by whether they make something pretty in 30 seconds. I would judge them by a harder question: Can they preserve intent from idea to build? If yes, they are not just creative tools. They start becoming workflow infrastructure. 👀 🔗 https://lnkd.in/gW4HsB5c #AI #DesignSystems #ProductDevelopment #Anthropic #Claude #AIArchitecture
-
Saeed Kasmani, Ph.D. shared thisPurpose-built AI is where the real enterprise value will show up first. Siemens just launched the Eigen Engineering Agent for industrial automation, and the interesting part is not that it can talk about engineering. It can work inside the actual engineering stack, with project context, standards, device relationships, PLC logic, HMI visualisation, and configuration tasks. That is a much more important product pattern than another generic copilot. In most engineering environments, the bottleneck is not access to an LLM. It is the translation layer between broad model output and the messy constraints of the real system. The more safety, context, and domain structure a workflow needs, the less useful a generic assistant becomes on its own. Siemens says the agent can complete workflows 2 to 5 times faster than manual alternatives, with up to 80% higher solution quality and 50% greater engineering efficiency. Even if those numbers soften outside vendor benchmarks, the direction still makes sense. The next wave of enterprise AI value probably will not come from one general-purpose assistant trying to do everything. It will come from narrower agents that are deeply wired into a domain, know the local context, and can produce work that is actually ready for review. That is where AI stops being interesting and starts becoming operational. ⚙️ 🔗 https://lnkd.in/gSC2iBmp #AI #IndustrialAI #AutomationEngineering #AIAgents #EnterpriseAISiemens launches the Eigen Engineering Agent, bringing purpose-built AI to industrial automation engineeringSiemens launches the Eigen Engineering Agent, bringing purpose-built AI to industrial automation engineering
-
Saeed Kasmani, Ph.D. shared thisOne of the more important AI infrastructure shifts right now is that inference is starting to get its own control plane. That is why Google Cloud talking about Inference Gateway alongside llm-d is more interesting than it first sounds. A lot of teams still think of inference as “deploy a model endpoint and scale it.” That view is getting too simple. Once workloads become more varied, the hard part is not only serving tokens. It is deciding: • Where traffic should go • Which model tier should answer • How to balance latency against cost • How to keep throughput high without wrecking tail latency • How to operate this inside a real platform, not just a benchmark setup That is why I think inference architecture is becoming its own discipline. The winning stack is not just the best model. It is the combination of: • Routing • Gateway behaviour • Batching strategy • Scheduler decisions • Observability • Failure handling In other words, the system around the model is becoming part of the product. That is a healthier direction for the industry. It moves the conversation away from leaderboard theatre and toward architecture that actually survives production. ⚙️ 🔗 https://lnkd.in/gaFitNwV #AI #AIInfrastructure #MLOps #Kubernetes #Inference #AIArchitecture #PlatformEngineering
-
Saeed Kasmani, Ph.D. shared thisMarketing AI often stops at recommendations. It can score a segment. Suggest a message. Maybe even generate the asset. Then the handoff starts. Adobe's CX Enterprise Coworker is interesting because it tries to cut through that handoff. Adobe says it works across Experience Platform, Real-Time CDP, Journey Optimiser, and Customer Journey Analytics, so a team can move from customer signal to next-best action in a single flow, with a person still supervising the decision. That is a much bigger shift than "AI for campaign content." If this category matures, the value will not come from nicer prompts. It will come from tighter decision loops. Who should we act on right now? What should happen next? Which channel makes sense? What needs approval? What changed after the intervention? Adobe says AEP now supports more than one trillion experiences a year. At that scale, speed matters, but sequence matters even more. The teams that win here will not be the ones with the flashiest model. They will be the ones who can connect customer context, decisioning, and activation without turning the stack into a bowl of spaghetti. 🎯 🔗 https://lnkd.in/g7JmtNZA #AI #CustomerExperience #MarTech #Decisioning #PersonalisationAdobe Unveils CX Enterprise Coworker to Build Agentic-Enabled Workflows for Customer Experience OrchestrationAdobe Unveils CX Enterprise Coworker to Build Agentic-Enabled Workflows for Customer Experience Orchestration
-
Saeed Kasmani, Ph.D. shared thisGitHub’s new Copilot AI Model Providers status component looks like a small ops update. I think it is a much bigger signal than that. Once your AI product can route across multiple models, reliability stops being a single-service metric. You need to know whether the issue is your product, your routing layer, or one provider behind it. That sounds operationally obvious, but a lot of teams still report AI availability as one blended number. That hides the real failure mode. GitHub is now separating provider disruptions from the Copilot service itself, publishing per-service uptime, and adding a degraded performance state instead of collapsing everything into "outage". That is a better way to run any serious multi-model platform. Why it matters: • buyers get a more honest view of platform risk • engineering teams can design cleaner fallbacks and escalation paths • product teams can stop pretending every model incident is the same incident As more AI systems adopt auto-routing, this kind of provider-level observability will become normal. The interesting competition is no longer just model quality. It is quality plus routing plus transparency when something breaks. ⚙️ 🔗 https://lnkd.in/gEjjNvsU #AI #AIArchitecture #MLOps #LLMOps #GitHubCopilot #AIProductsBringing more transparency to GitHub’s status pageBringing more transparency to GitHub’s status page
-
Saeed Kasmani, Ph.D. liked thisSaeed Kasmani, Ph.D. liked thisWhat a day at the Microsoft AI Tour for Partners day in Sydney on Monday. 700+ attendees from 300+ Microsoft partner organisations came together to explore how AI is reshaping what's possible across our partner ecosystem — and the energy on the day was amazing. For our SDC/ISV audience, we had deep and informative sessions on Build to Bill for AI Solutions, Marketplace, Benefits & Incentives, Certified Software Designation and closed out with a lively panel discussion around partnering with Microsoft to deliver success. The conversations with partners about turning AI ambition into real customer outcomes were definitely a highlight. Thank you to the presenters on the day Madeline Beattie, Kristyn Maddox, Olivia Nix, Danny Chua, Sinéad Fitzgerald and David Wright. A huge thank you to our partner panelists who joined us Cathy Conroy, Jaimie Tilbrook, Michael Horin and Rod O'Rourke. Partner day wrapped, and now looking forward to an amazing customer-facing AI Tour event tomorrow at the Sydney ICC with over 6,000 people! #MicrosoftAITour #AIforPartners #ISV #AI #Microsoft Vicki Batka MAICD Elena Reinhardt Glen Strike James Paek Madeline Beattie Ryan Tarak Simon Schultheiss Stephanie Neergaard Marc Havercroft Matt F. Meera Bangi
-
Saeed Kasmani, Ph.D. liked thisDoes your organisation use TechnologyOne? There is now a direct VAPAR integration that syncs your CCTV inspections & defects between systems. This means the data moves exactly where you need it - automatically! You can check out the full details in the webinar linked below, or get in touch with me and I can run you through it one-on-one.Saeed Kasmani, Ph.D. liked thisCouncils and utilities are under pressure to act faster on inspection data, not just analyse it. Access the latest webinar co-presented with TechnologyOne to see how a customer-led integration is helping utilities move from CCTV footage to prioritised defects and actionable work orders, without adding complexity. ✔ Faster triaging of defects from CCTV inspections ✔ Seamless creation of work orders for high-risk assets ✔ AI-powered condition assessments to support decisions ✔ Built based on real customer workflows If improving response times and turning data into action is a priority, this session is for you. 👉 Watch the webinar here: https://lnkd.in/gQYSbBVX #AssetManagement #WaterUtilities #Wastewater #Stormwater #AI #DigitalTransformation #Infrastructure #CapitalPlanning #SmartUtilities #GovTech #TechnologyOne #VAPAR #sewer #drainage
Experience & Education
-
IBM
View Saeed’s full experience
See their title, tenure and more.
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
Licenses & Certifications
Publications
-
A-CCNN: adaptive ccnn for density estimation and crowd counting
ICIP 2018/ IEEE
Crowd counting, for estimating the number of people in a crowd using vision-based computer techniques, has attracted much interest in the research community. Although many attempts have been reported, real-world problems, such as huge variation in subjects' sizes in images and serious occlusion among people, make it still a challenging problem. In this paper, we propose an Adaptive Counting Convolutional Neural Network (A-CCNN) and consider the scale variation of objects in a frame adaptively…
Crowd counting, for estimating the number of people in a crowd using vision-based computer techniques, has attracted much interest in the research community. Although many attempts have been reported, real-world problems, such as huge variation in subjects' sizes in images and serious occlusion among people, make it still a challenging problem. In this paper, we propose an Adaptive Counting Convolutional Neural Network (A-CCNN) and consider the scale variation of objects in a frame adaptively so as to improve the accuracy of counting. Our method takes advantages of contextual information to provide more accurate and adaptive density maps and crowd counting in a scene. Extensively experimental evaluation is conducted using different benchmark datasets for object-counting and shows that the proposed approach is effective and outperforms state-of-the-art approaches.
Other authors -
-
Structured Inhomogeneous Density Map Learning for Crowd Counting
arXiv preprint arXiv:1801.06642
In this paper, we aim at tackling the problem of crowd counting in extremely high-density scenes, which contain hundreds, or even thousands of people. We begin by a comprehensive analysis of the most widely used density map-based methods, and demonstrate how easily existing methods are affected by the inhomogeneous density distribution problem, e.g., causing them to be sensitive to outliers, or be hard to optimized. We then present an extremely simple solution to the inhomogeneous density…
In this paper, we aim at tackling the problem of crowd counting in extremely high-density scenes, which contain hundreds, or even thousands of people. We begin by a comprehensive analysis of the most widely used density map-based methods, and demonstrate how easily existing methods are affected by the inhomogeneous density distribution problem, e.g., causing them to be sensitive to outliers, or be hard to optimized. We then present an extremely simple solution to the inhomogeneous density distribution problem, which can be intuitively summarized as extending the density map from 2D to 3D, with the extra dimension implicitly indicating the density level. Such a solution can be implemented by a single Density-Aware Network, which is not only easy to train, but also can achieve the state-of-art performance on various challenging datasets.
Other authors -
-
A new multi-objective cluster ensemble based on modularity maximization
Journal of Engineering Research (JER)
Conventional clustering algorithms utilize only one single criterion that may not conform to diverse shapes of the underlying clusters. But in this paper, we hire two important criteria and propose a new multi-objective cluster ensemble model to empower finding clusters of different types. The first criteria are the well-known sum of squared error. The second criterion is modularity which is originally a measure of evaluating communities in social networks. We maximize modularity as a consensus…
Conventional clustering algorithms utilize only one single criterion that may not conform to diverse shapes of the underlying clusters. But in this paper, we hire two important criteria and propose a new multi-objective cluster ensemble model to empower finding clusters of different types. The first criteria are the well-known sum of squared error. The second criterion is modularity which is originally a measure of evaluating communities in social networks. We maximize modularity as a consensus function of cluster ensemble. In order to further improvement, we also modify Non Dominant Sorting Genetic Algorithm (NSGAII) and propose a specialized crossover operator for it. Experimental results over seven UCI real data sets show that the proposed method significantly outperforms other clustering methods.
Other authors -
-
Intrusion Detection Based on Joint of K-Means and KNN
JCIT (Journal of Convergence Information Technology)
-
Availability challenge of cloud system under DDOS attack
Indian Journal of Science and Technology
Abstract Cloud computing is a distributed architecture that has shared resources, software,
and information is provided to computers and other devices on a scalable platform and
demand. Availability of the cloud services is one of the key security issues in it. Distributed
Denial of Service (DDoS) is an attack that threats the availability of the cloud services. In this
paper, effect of the DDoS attack on the cloud is investigated. Therefore, a model for attack
based on the…Abstract Cloud computing is a distributed architecture that has shared resources, software,
and information is provided to computers and other devices on a scalable platform and
demand. Availability of the cloud services is one of the key security issues in it. Distributed
Denial of Service (DDoS) is an attack that threats the availability of the cloud services. In this
paper, effect of the DDoS attack on the cloud is investigated. Therefore, a model for attack
based on the DDoS is designed, and then we simulate a cloud system on the ...Other authors -
-
A new robust digital image watermarking technique based on joint DWT-DCT transformation
Third International Conference on Convergence and Hybrid Information Technology, 2008. ICCIT'08.
Abstract In this paper, a new robust digital image watermarking algorithm based on Joint
DWT-DCT Transformation is proposed. The imperceptibility is provided as well as higher
robustness against common signal processing attacks. A binary watermarked image is
embedded in certain sub-bands of a 3-level DWT transformed of a host image. Then, DCT
transform of each selected DWT sub-band is computed and the PN-sequences of the
watermark bits are embedded in the coefficients of…
Abstract In this paper, a new robust digital image watermarking algorithm based on Joint
DWT-DCT Transformation is proposed. The imperceptibility is provided as well as higher
robustness against common signal processing attacks. A binary watermarked image is
embedded in certain sub-bands of a 3-level DWT transformed of a host image. Then, DCT
transform of each selected DWT sub-band is computed and the PN-sequences of the
watermark bits are embedded in the coefficients of the corresponding DCT middle ...Other authors -
-
A New Method for Improving the Performance of K Nearest Neighbor using Clustering Technique.
Journal of Convergence Information Technology - JCIT
Abstract In this paper, a new classification method is presented which uses clustering
techniques to augment the performance of K-Nearest Neighbor algorithm. This new method
is called Nearest Cluster approach, NC. In this algorithm the neighbor samples are ...Other authors -
-
Robust Digital Image Watermarking Based on Joint DWT-DCT.
International Journal of Digital Content Technology and its Applications (JDCTA)
-
A Pre-Filtering Method to Improve Watermark Detection Rate in DCT Based Watermarking
International Arab Journal of Information Technology (IAJIT)
Abstract: In image processing pre-processing is used for preparing or improving
performance of operations. In order to improve performance of extraction algorithms in
Discrete Cosine Transform (DCT) based watermarking method, a new prefiltering method ...Other authors -
-
Effect of Locations of Using High Boost Filtering on the Watermark Recovery in Spatial Domain Watermarking
Indian Journal of Science & Technology
Abstract Generally, High Boost filter is used to emphasize high frequency components
representing the image details without eliminating low frequency components representing
the basic form of the signal. The dissimilarity between the watermark and unwatermarked
parts of the image are increased by this filter. Thus, watermark could be recovered
significantly better by recovery algorithm. In this paper, a comparison is taken place between
the effects of different places of…
Abstract Generally, High Boost filter is used to emphasize high frequency components
representing the image details without eliminating low frequency components representing
the basic form of the signal. The dissimilarity between the watermark and unwatermarked
parts of the image are increased by this filter. Thus, watermark could be recovered
significantly better by recovery algorithm. In this paper, a comparison is taken place between
the effects of different places of preforming High Boost filter on the reliability of the ...Other authors -
Courses
-
Advanced AI
-
-
Computer Vision
-
-
Digital Image processing
-
-
Digital Signal Processing
-
-
Machine Learning
-
-
Neural Network
-
-
Statistical Pattern Recognition
-
-
Symbolic Processing - Logic, Theorem proving & Term rewriting
-
Projects
-
Towards Development of Automated Diagnostic Tools for Pneumoconiosis Detection from Chest Radiographs
- Present
In this project, we have been working on the application of Deep Learning on accurately processing the medical image to find potential signs of certain diseases. In order to develop a program for this purpose, we have been using python, Nvidia digits and Keras Deep Learning packages in Ubuntu Linux OS.
-
Rail Manufacturing CRC and Sydney Trains with UTS Project about human activity analysis
- Present
In this project, we have been utilizing Deep Learning models to analyse CCTV camera. The purpose of this project is to extract valuable information from current infrastructure and provide better and more secure services for customers. I have been using python, Caffe and Tensorflow Deep Learning to do this project.
-
Developed an end-to-end demo for Consumer complaint classification based on NLP
-
-
Developed an end to end AI-based Solution for detection and localising defect in the sewer and stormwater pipe
-
-
Duplicate defect detection by Machine Learning model
-
-
Water level estimation for sewer and storm water pipe
-
-
Face matching and profiling
-
-
Person and object detection and localisation
-
Honors & Awards
-
Published over 25 papers in well-known journals and conferences
-
-
Reached over 430 citations for the published papers
-
-
Winner of ICIP 2019 VIP cup in Taipei
IEEE, ICIP 2019
#UTS and #PolyU joint team win the Championship of 2019 VIP cup in Taipei. I was the supervisor of the UTS team.
The final was held earlier yesterday at #ICIP2019 ( ICIP is the world's largest and most comprehensive technical conference focused on image and video processing and computer vision). -
One-Off Scholarship
Graduate Research School
-
Data61 Top-up Scholarship
CSIRO \ Data61
Data61, the digital powerhouse formed by the merging of CSIRO's digital productivity business unit and National ICT Australia (NICTA), offers a world-class Ph.D. experience to eligible Ph.D. candidates working in data-related disciplines such as Analytics, Cyber-Physical Systems, Software, and Computational Systems and Decision Sciences.
-
International Research Scholarship (IRS) - International Research Scholarships
UTS
The UTS International Research Scholarship (UTS IRS) is provided by the University of Technology, Sydney (UTS) as part of its long term commitment to internationalization with a particular view to enhancing its international links and profile in research. It is aimed at attracting high-quality international students to work in areas of research strength at UTS.
-
UTS President's Scholarship
UTS
The UTS President's scholarship (UTSP) is provided by the University of Technology Sydney to international Higher Degree Research students who demonstrate exceptional research potential. UTSPs are provided to assist with students' general living costs. The UTSP scholarship is not transferable to another institution.
-
Best Researcher of Islamic Azad University
Azad University
I have been selected as the Best Researcher of Islamic Azad University, Ramsar Branch from 2015. I received this award due to my excellent achievements for completing two research projects and publishing several journal papers.
-
Best Researcher of Islamic Azad University
Azad University
I have been selected as the Best Researcher of Islamic Azad University, Ramsar Branch from 2013. I received this award due to my excellent achievements for completing two research projects and publishing several journal papers.
-
"First Class Honour" Graduate in BSc
-
Graduated with “First Class Honour” and 4rd rank over more than 50 students (BSc)
-
Ranked Top 3% in National University Entrance
ExamNational Organisation for Educational Testing
Ranked Top 3% (321 over more than 12,000 graduate computer science and IT applicants) in National Entrance Exam
p.s. "The National University Entrance Exam is a very competitive standardized test so that every applicant needs to pass it to gain admission to higher education in Iran".
(Source: http://en.wikipedia.org/wiki/Iranian_University_Entrance_Exam)
Languages
-
English
Full professional proficiency
-
Persian
Native or bilingual proficiency
Recommendations received
11 people have recommended Saeed
Join now to viewExplore more posts
Explore collaborative articles
We’re unlocking community knowledge in a new way. Experts add insights directly into each article, started with the help of AI.
Explore More