Articles by Tom
-
Grey Jean Technologies Launches AI-Powered Personalization for Retailers
Grey Jean Technologies Launches AI-Powered Personalization for Retailers
http://www.businesswire.
Activity
7K followers
-
Tom Lorenc shared thisArchitecting enterprise AI systems using multi-agent reasoning frameworks, hybrid retrieval architectures, and knowledge graphs that combine structured enterprise data with unstructured document intelligence. I created the Closed-Loop Agentic AI Systems (CLAAS) governance framework, which provides a structured architecture for deploying large language models in enterprise environments with traceability, evaluation pipelines, and human-in-the-loop oversight.
-
Tom Lorenc shared thisaskUH Find-a-Doctor — Multi-Agent AI Provider Search Platform | University Hospitals Designed and deployed an 8-agent AI system on Azure AI Foundry (GPT-4.1) that enables patients to find healthcare providers through natural language conversation, replacing a rigid intake form with NLP-first intent inference. Built a 3-layer agent pipeline — User Agent (symptom-to-specialty classification), Route Agent (specialty dispatcher), and 6 Specialty Agents — each querying dedicated Azure AI Search indexes across 3,371 provider-location documents with a tiered geo-search strategy (Top 5 → 10mi → 25mi → 40mi radius expansion). Developed a standalone Azure Functions distance bucketing service using Haversine calculation over a 2,755-ZIP centroid table covering Ohio and 5 bordering states. Implemented emergency safety protocols (911 escalation for cardiac symptoms), sensitivity-aware responses for reproductive health, and a zero-downtime deployment pipeline. Upgraded schema from 15 to 27 fields with Edm.GeographyPoint geo-search while maintaining full backward compatibility across agent names, workflows, and indexes.
-
Tom Lorenc shared thisMars Requirements Extractor Summary This is a Python system that automatically extracts vendor compliance requirements from retail routing guide PDFs (Kroger, CVS, Walmart) using GPT-5.1 via the MAX Gateway. The Problem: Logistics teams manually read 30-50 page PDFs with 100+ compliance rules, miss requirements, and pay millions in penalty fees. The Solution: LLM-powered extraction + vector database for semantic search. Pipeline Flow Ingest — Extract text from PDF using PyMuPDF (PyPDF2 fallback) Deduplicate — Remove duplicate pages via SHA-256 hashing Chunk — Split into ~14K character chunks respecting page boundaries Extract — Send to GPT-5.1 with structured prompts → JSON output Validate/Repair — Schema validation with LLM repair loop if needed Global Dedupe — Token Jaccard similarity (0.92 threshold) across chunks Output — 21-column CSV + PostgreSQL with pgvector embeddings Key Technical Components ComponentTechPurposeExtractionGPT-5.1PDF → structured JSONEmbeddingstext-embedding-3-smallSemantic vectorsStoragePostgreSQL + pgvectorVector similarity searchDedupUnion-find + scoringRemove duplicate rules Why pgvector? Semantic search finds related requirements even with different wording—"late" matches "tardy," "delayed," and "missed appointment."
-
Tom Lorenc shared thisBAIN CAPITAL - Multi Agent LangGraph Orchestration, LangChain agents for Value Chain, Market Map
-
Tom Lorenc shared thisBuilt an enterprise-grade Agentic Generative AI Cancer Expert Chat Bot leveraging AWS Bedrock's Claude and Llama models, enhanced with LangGraph multi-agent coordination for comprehensive cancer detection and clinical trial matching workflows.
-
Tom Lorenc shared thisMore amazing work done by an incredible groupHarmonEyes Introduces Theia™ – The Ultimate Eye-Tracking Deployment EngineHarmonEyes Introduces Theia™ – The Ultimate Eye-Tracking Deployment Engine
-
Tom Lorenc shared thisHarmonEyes Introduces Theia™ – The Ultimate Eye-Tracking Deployment EngineHarmonEyes Introduces Theia™ – The Ultimate Eye-Tracking Deployment Engine
-
Tom Lorenc liked thisTom Lorenc liked thisMost LLM agents hit a ceiling after deployment. The model is frozen, so it cannot learn from what actually happens in production. The only way to improve is retraining or prompt tweaks, both of which are slow, expensive, and brittle. The paper Memento-Skills: Let Agents Design Agents changes that assumption. It keeps the model fixed and moves learning into a separate layer. Instead of updating weights, the system updates a growing library of skills. In this setup, memory is not just logs or past conversations. It is a collection of reusable skills. Each skill is something the agent can actually execute, including prompts, code, and instructions. When a task comes in, the agent picks a skill, runs it, checks the result, and then updates that skill based on what went right or wrong. Learning happens by improving these skills over time. The loop is simple in practice. Retrieve a skill, execute it, evaluate the result, then rewrite the skill. Over time, the agent builds better and more reliable ways of solving tasks without ever changing the underlying model. A key design decision is how the agent picks a skill. Most systems use embedding similarity, which often fails in practice. Similar wording does not mean the same solution. This paper instead trains the router to select skills based on whether they actually work, using a reinforcement learning signal tied to execution success. That aligns retrieval with outcomes, not just text similarity. The results reflect this shift. On GAIA, adding this skill evolution layer improves accuracy from 52.3% to 66.0%. On HLE, performance more than doubles compared to a system that retrieves and executes but does not update its skills. At the same time, there is a clear limitation. This only works well when tasks share structure. If tasks are too diverse, many learned skills are never reused. This is a quite important development because core system is no longer just a model with tools. It is a system that maintains and improves a skill library. That means treating skills as first-class artifacts. They need versioning, testing, performance tracking, and controlled updates. It also means evaluating three separate things: whether the right skill was selected, whether the task was solved, and whether the execution was safe. What this paper makes explicit is where learning should live. The model is no longer the thing that improves. It becomes a stable execution engine. All adaptation moves into a separate layer that you can observe, test, and update directly. In practice, that means the system improves by refining how tasks are solved, not by changing how the model thinks. Every failure updates a skill. Every success reinforces one. Over time, the agent builds a library of working solutions instead of relying on a static prompt or a retrained model. https://lnkd.in/eEj-Bbj9 https://lnkd.in/eK_gEB_c
-
Tom Lorenc liked thisTom Lorenc liked thisWhat if the noise can be turned into signal? Once in a while you come across a paper where the idea is so refreshing. HiFiGaze (CMU, CHI '26) is one such work. In eye tracking, pupil corneal reflection from fixed IR LEDs is used for gaze estimation. HiFiGaze bootstraps that same geometric intuition using the phone screen itself as the illumination source, with screen content knowledge as the disambiguation mechanism (thanks to 4K selfie cameras). The device already knows what's on its own screen, so content variability is no longer a problem. No IR LEDs, no calibration, no extra hardware. That's a genuinely creative inversion of the problem! 👏 arxiv.org/abs/2603.19588 #EyeTracking #GazeEstimation #ComputerVision #HCI #CHI2026
-
Tom Lorenc liked thisTom Lorenc liked thisGoogle's AMIE (Articulate Medical Intelligence Explorer, an LLM-based conversational diagnostic system) just completed its first prospective clinical trial with real patients, and the results suggest conversational AI in primary care is closer to reality than most people think. 1️⃣ 100 patients at Beth Israel Deaconess chatted with AMIE (built on Gemini 2.5) up to 5 days before their urgent care visit. Transcripts and summaries were sent to their PCP. 2️⃣ Zero safety stops across all interactions, with board-certified internists monitoring every conversation in real time. 3️⃣ AMIE's differential included the final diagnosis (per 8-week chart review) in 90% of cases. Top-3 accuracy: 75%. Top-1: 56%. 4️⃣ Blinded evaluators found no significant difference between AMIE and PCPs for differential quality (p = 0.6), management plan appropriateness (p = 0.1), or safety (p = 1.0). 5️⃣ PCPs scored higher on practicality (p = 0.003) and cost effectiveness (p = 0.004), likely because AMIE had no EHR access, no physical exam, and no prior patient context. 6️⃣ Patient attitudes toward AI improved significantly after using AMIE (p < 0.001) and stayed elevated after seeing their PCP. 7️⃣ 75% of PCPs who reviewed the transcript found it helpful. 64% trusted the information. PCPs compared it to a third-year medical student's intake. 8️⃣ Patients arrived more organized, shifting visits from data gathering to collaborative decision-making. 9️⃣ Real-world friction: 7% couldn't complete the study due to device issues, only 73% of PCPs reviewed the transcript pre-visit because AMIE wasn't EHR-integrated. 🔟 Single-arm feasibility study, one center, N=100, excluded pregnant and mental health patients. Promising but far from definitive. ✍🏻 Peter Brodeur, MD, MA, Jacob M. Koshy, Anil Palepu, Khaled Saab, Ava Homiar, Roma Ruparel, MPH, Charles Wu, Ryutaro Tanno, Joseph Xu, Amy Wang, David Stutz, Hannah Mulvey Ferrera, David Barrett, Lindsey Crowley, Jihyeon Lee, Spencer E. Rittner, Ellery Wulczyn, Selena K. Zhang, Elahe Vedadi, Christine G. Kohn, Kavita Kulkarni, Vinay Kadiyala, S. Sara Mahdavi, Wendy Du, Jessica M. Williams, David Feinbloom, Renée Wong, Tao Tu, Petar Sirkovic, Alessio Orlandi, Christopher Semturs, Yun Liu, Juraj Gottweis, Dale R. Webster, Joëlle Barral, Katherine Chou, Pushmeet Kohli, Avinatan Hassidim, Yossi Matias, James Manyika, Rob Fields, Jonathan X. Li, Marc L. Cohen, Vivek Natarajan, Mike Schaekermann, Alan Karthikesalingam MD PhD, Adam Rodman. A prospective clinical feasibility study of a conversational diagnostic AI in an ambulatory primary care clinic. arXiv. 2026. DOI: 2603.08448v2
-
Tom Lorenc liked thisTom Lorenc liked thisThere’s a lot of focus right now on what AI can do, of course. Less discussion on what actually makes those capabilities valuable inside a product. In practice, a few things seem to matter more than anything else: - latency - clarity of outputs - how easily something fits into an existing workflow We’ve spent a lot of time improving system performance and simplifying how data moves through the platform, and it’s interesting how often that unlocks more value than adding something new. AI increases potential. System performance and product clarity determine whether that potential shows up in real usage. Feels like the most effective teams right now are paying equal attention to both.
-
Tom Lorenc liked thisTom Lorenc liked thisMeituan just released a 560-billion-parameter open-source MoE model trained specifically for formal theorem proving. LongCat-Flash-Prover treats mathematical reasoning as an interactive dialogue with Lean4. Rather than generating proofs in a single pass, it decomposes the task into three agentic capabilities: auto-formalization, sketching, and proving. The model actively queries the compiler to verify syntax and semantic consistency, refining its approach based on real-time environmental feedback. The results significantly advance the state of the art for open-weights models. With only 72 inference attempts per problem, it achieves a 97.1% pass rate on MiniF2F-Test. On more challenging benchmarks, it solves 70.8% of ProverBench and 41.5% of the notoriously difficult PutnamBench problems. To stabilize training on these long-horizon tasks, the team developed HisPO (Hierarchical Importance Sampling Policy Optimization), which applies gradient masking to handle discrepancies between training and inference engines. They also incorporated legality detection to prevent reward hacking during reinforcement learning. Model weights are available now under the MIT License. https://lnkd.in/eChDhmEZ https://lnkd.in/eacSVJSQ https://lnkd.in/e2ikiFip
-
Tom Lorenc liked thisTom Lorenc liked thisMisleading AI explanations lowered novice students’ diagnostic accuracy and broke confidence calibration, while correct explanations did not raise accuracy over no explanations. 1️⃣ Three‑arm randomized trial with 111 junior medical students compared no explanations, correct AI explanations, and misleading AI explanations. 2️⃣ Misleading AI cut accuracy to 9.2% vs 21.0% in control. 3️⃣ Correct AI yielded 22.6% accuracy and no significant improvement over control. 4️⃣ Any AI explanation inflated confidence compared with control, regardless of being correct or misleading. 5️⃣ Confidence calibration differed by group: present with correct explanations, absent with misleading explanations and control. 6️⃣ Misleading AI steered choices toward a specific wrong option; 70.27% of incorrect answers were targeted commission errors vs 24.65% in control. 7️⃣ Harm varied by item; for the most persuasive items, accuracy in the misleading group dropped to 6.8%. 8️⃣ Groups were demographically balanced at baseline, supporting causal interpretation of group differences. 9️⃣ Educational takeaway: novice learners need AI‑auditing drills and critical appraisal workflows to prevent high‑confidence errors. 🔟 Limitations include 100% correct vs 100% incorrect explanations, 25 items, single cultural context, and novice-only sample, so generalizability may be limited. ✍🏻 Da Teng, Lihua Tan, Qiyuan Cao, Yanwei Xia, Na Zhang, Jiantao Li, Dan Zhao. Impact of AI misinformation on diagnostic accuracy and confidence calibration in novice medical students. npj Digital Medicine. 2026. DOI: 10.1038/s41746-026-02547-z
Experience & Education
-
HarmonEyes - Healthcare AI
View Tom’s full experience
See their title, tenure and more.
or
Patents
-
Decentralized Distributed Machine Learning
Filed US Application No. 62/701,238
See patentAspects of the invention provide a decentralized, distributed machine learning
environment. In a particular embodiment, a machine learning solution is implemented
using a blockchain. An embodiment distributes both the data and the machine learning
models to multiple computer systems in the environment. Each computer system
calculates the back propagation weights. Blockchain technology can be used to provide
serial, secure communications between the computer systems (e.g.…Aspects of the invention provide a decentralized, distributed machine learning
environment. In a particular embodiment, a machine learning solution is implemented
using a blockchain. An embodiment distributes both the data and the machine learning
models to multiple computer systems in the environment. Each computer system
calculates the back propagation weights. Blockchain technology can be used to provide
serial, secure communications between the computer systems (e.g., calculating a nonce
used to encrypt communications), propagation of changes to other computer systems,
Courses
-
computer science
-
Projects
-
Visible World - targeted video streaming
-
See projectArchitected, designed and hands on developed a video ad web application in php/Java that allows a user to create custom video ads online, target that video to any web site, place that video in rotation and allows for advertising in stream and preroll. System built using in php, Java J2EE WebSphere using SOA, Spring, Struts, Hibernate
-
POC for Oncology Decision Support System: FHIR-based LLM RAG Architecture
-
This POC integrates NCCN oncology guidelines with FHIR standards and LLM technology to create an intelligent clinical decision support system. The system leverages RAG (Retrieval Augmented Generation) with a React frontend, OpenAI LLM backend, and Pinecone vector database with advanced re-ranking techniques.
-
RAG reranking agent
-
RAG reranking agent
-
ZML AI_ML RAG Multi Agent Graph Architecture
-
ZML AI_ML RAG Multi Agent Graph Architecture
Languages
-
English
Native or bilingual proficiency
Recommendations received
-
LinkedIn User
6 people have recommended Tom
Join now to viewOther similar profiles
Explore more posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content