Connected project evidence
Explore My AI Impact
17 projects across 9 impact domains, connected through research, production AI, computer vision, intelligent systems, reliability, and optimization.
Explore Impact Domains
AI engineering portfolio
Research systems & production AI for real-world impact
Explore technical case studies through the problem they solve, the capabilities they use, or the systems they connect.
- Projects
- 17
- Capabilities
- 15
- Industries
- 7
- Impact domains
- 9
Project discovery mode
17 projects
AND across filters · OR within capabilities
DeltaCert-Agent
A change-impact-aware security recertification framework for evolving tool-using LLM agents that maps configuration changes to affected assurance claims, selects scoped tests, executes risk-triggered sentinels, and escalates to broader recertification when impact cannot be bounded safely.
0.7502 recall regression detection vs 0.5501 equal-budget random
Controlled Evaluation 31,396 evidence rows · four local models · five repetitions
How measured Research · Data basis Architecture
Capabilities
- Security
- Agentic AI
- More capabilities
Why it matters
It makes security recertification proportional to bounded change impact while retaining a conservative path to full-suite testing whenever selective evidence is unsafe.
Research
CARE-SIU
A research framework for temporal and multimodal social-interaction understanding that combines R3D-18 video encoding, reliability-aware fusion, calibration, explainability, synthetic-to-real evaluation, and leakage-resistant experimentation for safety-critical environments.
0.8134 ± 0.0121 macro-F1
Benchmark RWF-2000 real-video benchmark · five seeds
How measured Research · Data basis Architecture
Capabilities
- Multimodal AI
- Computer Vision
- More capabilities
Why it matters
Safety-critical interaction understanding must expose evidence quality, uncertainty, calibration, missing modalities, and evaluation leakage instead of returning an unexplained label.
Research
IncidentGraph
IncidentGraph is a research prototype for uncertainty-aware multi-camera incident reconstruction that preserves evidence provenance, contradictory hypotheses, alternatives, sensor gaps, confidence, and cross-camera associations in a typed graph.
0.923 ± 0.069 diagnostic graph score
Controlled Evaluation 60 synthetic multimodal incidents
How measured Research · Data basis Architecture
Capabilities
- Computer Vision
- Multimodal AI
- More capabilities
Why it matters
Evidence fusion should preserve provenance, contradictions, alternatives, and sensor gaps so investigators can distinguish 'not observed' from 'did not occur' and audit how a reconstruction was formed.
Healthcare
MedClaim Sentinel
A local-first medical insurance claim review platform that reads multilingual claim documents, detects repeated claims and invoices, reconciles amounts, performs conservative diagnosis–medicine–bill–investigation matching, and supports evidence-grounded reviewer chat.
Prototype · Video Architecture
Capabilities
- Document Intelligence
- Multimodal AI
- More capabilities
Why it matters
Claims reviewers must reconcile diagnoses, medicines, invoices, duplicate claims, and billed amounts across fragmented documents, making manual review slow and inconsistent.
Manufacturing
Engineering Work Order P&L Analytics
A synthetic-data-based analytics platform for estimating work-order revenue, tracking delivery-center and field-operations labor, monitoring billing and other costs, calculating profit and loss, and identifying margin risk across the engineering work-order lifecycle.
Prototype · Data basis Architecture
Capabilities
- Data Analytics
- Workflow Automation
Why it matters
Without unified P&L per work order, margin leaks hide inside labor and billing noise until quarter-end.
Telecom
Telecom Copper Reclamation AI
A decision-support and workflow-modernization platform for telecom copper reclamation that parses ACR, CAPR, and CPR reports, normalizes cable-pair evidence, generates conservative review recommendations, models With-SOW and Without-SOW processes, and provides a FastAPI and React/Vite foundation for governed automation.
Prototype · Data basis Architecture
Capabilities
- Document Intelligence
- Workflow Automation
- More capabilities
Why it matters
Reclamation decisions worth millions hinge on evidence scattered across ACR, CAPR, and CPR reports that no one reads end-to-end.
Security
LightDID-ZKP
LightDID-ZKP is a research framework introducing CAPS-ZK, a policy- and resource-aware selector for BBS and AnonCreds privacy-preserving verifiable presentations.
2,405-byte VP BBS presentation size at 64 attributes
Controlled Evaluation 50 measured runs per configuration
How measured Research · Paper Architecture
Capabilities
- Security
Why it matters
Choosing the wrong proof scheme leaks more identity data or burns more compute than the policy requires.
ASHU Mentor AI Studio
ASHU Mentor AI Studio is a local-first AI mentor platform that connects resume-aware interviews, JD-based evaluation, adaptive training, evidence capture, consent-based voice generation, and digital-human lecture rendering into one complete learning workflow.
Live / Deployed · Live Demo · Architecture Production
Capabilities
- Workflow Automation
- Generative AI
- More capabilities
Why it matters
Career preparation is fragmented across separate interview, resume, evaluation, coaching, voice, and digital-human training tools.
Security
AegisFlow
An agentic DevSecOps cockpit for Azure Function and Python API repositories that validates CI/CD readiness, runs quality and security gates, explains failures, proposes human-approved fixes, and generates downloadable audit-ready evidence packs.
Live / Deployed · Live Demo · Architecture Production
Capabilities
- Security
- Agentic AI
- More capabilities
Why it matters
It turns fragmented CI/CD failures and scanner logs into one evidence-backed, human-governed readiness decision for developers, reviewers, security teams, and auditors.
Energy
Predictive & Preventive Maintenance for Generator Reliability
A full-stack analytics and ML dashboard that measures PM effectiveness, forecasts 14/30-day generator failure risk, and recommends prioritized maintenance actions.
Live / Deployed · Live Demo · Architecture Production
Capabilities
- Predictive Analytics
- MLOps
- More capabilities
Why it matters
It links maintenance activity to subsequent failures so teams can test whether PM schedules work, anticipate 14/30-day risk, and turn predictions into prioritized operational actions.
Logistics
Execution-Aware Agentic VRP Solver and Benchmark Studio
An end-to-end execution-aware Vehicle Routing Problem platform with 8 solver backends, 18 operational scenarios, and an Agentic AI layer powered by Google Gemini 2.0 Flash.
150/150 orders assigned with 0 late minutes and 0 overtime
Synthetic Evaluation 18 fixed scenarios · 8 solver backends
How measured Live / Deployed · Live Demo · Architecture Production
Capabilities
- Optimization
- Agentic AI
- More capabilities
Why it matters
Operational routing requires preserving work already in progress and balancing feasibility, lateness, overtime, distance, and runtime rather than optimizing distance alone.
End-to-End YOLO Key Detection System
End-to-end key detection using YOLO, FastAPI, Docker, Hugging Face Spaces, and Azure Container Apps for production-ready inference deployment.
Live / Deployed · Live Demo · Architecture Production
Capabilities
- Computer Vision
- Cloud Deployment
- More capabilities
Why it matters
It covers the operational gap between a trained detector and a servable, testable, containerized, cloud-deployable inference system.
Telecom
Pole Validation AI
AI-powered pole validation using GIS, imagery, OCR, and multi-system data reconciliation.
0.054 mAP@0.5 current detector performance
Validation Set Published training snapshot · 100 epochs · 640px images
How measured Prototype · Data basis Video
Capabilities
- Document Intelligence
- Computer Vision
Why it matters
Pole decisions affect cost, schedule, ownership, and safety, so visual detection must be reconciled with GIS and operational evidence rather than treated as a standalone model output.
Research
DACR-Q
A training-free inference framework that compensates INT4 quantization error with dynamic low-rank residual correction for memory-constrained LLM deployment.
Research · Data basis Architecture
Capabilities
- Generative AI
- Edge Deployment
Why it matters
It explores whether a lightweight input-conditioned correction can retain quantized-inference efficiency while recovering expressiveness without a retraining-heavy adaptation workflow.
Research
Vehicle-Scale LLMs
A memory-efficient in-vehicle LLM inference pipeline combining INT4 quantization with low-rank residual compensation for constrained edge hardware.
Research · Data basis Architecture
Capabilities
- Generative AI
- Edge Deployment
Why it matters
It demonstrates the complete mechanics of combining aggressive quantization with low-rank residual compensation before investing in real-model and embedded-hardware deployment.
Autonomous Microservice Composition
An agentic MCP control plane that converts natural-language intent into executable service DAGs using registry metadata, schema retrieval, orchestration, retries, and telemetry.
Prototype · Data basis Architecture
Capabilities
- Workflow Automation
- Agentic AI
- More capabilities
Why it matters
It shows how natural-language planning can be constrained by typed service knowledge and executed as an inspectable workflow graph instead of an opaque chain of model calls.
MCP 2.0
A multi-agent platform for orchestrating tools, memory, context, knowledge, and distributed workflows over gRPC and Protocol Buffers.
Prototype · Data basis Architecture
Capabilities
- Workflow Automation
- Agentic AI
- More capabilities
Why it matters
It treats multi-agent tool access, shared context, discovery, authorization, events, and delegation as governed infrastructure with inspectable contracts instead of ad hoc prompt integrations.
What problem are you trying to solve?
Explore relevant AI projects by business problem, technical challenge or application domain.
Computer Vision
Inspection, detection, tracking and video-intelligence systems.
5 projects
Technologies used
Similar use cases
Pole Validation AI
AI-powered pole validation using GIS, imagery, OCR, and multi-system data reconciliation.
Computer Vision Shared focus: Computer Vision · OCR · Pole Validation.
MedClaim Sentinel
A local-first medical insurance claim review platform that reads multilingual claim documents, detects repeated claims and invoices, reconciles amounts, performs conservative diagnosis–medicine–bill–investigation matching, and supports evidence-grounded reviewer chat.
Computer Vision Shared focus: OCR.
IncidentGraph
IncidentGraph is a research prototype for uncertainty-aware multi-camera incident reconstruction that preserves evidence provenance, contradictory hypotheses, alternatives, sensor gaps, confidence, and cross-camera associations in a typed graph.
Computer Vision Shared focus: Computer Vision.
End-to-End YOLO Key Detection System
End-to-end key detection using YOLO, FastAPI, Docker, Hugging Face Spaces, and Azure Container Apps for production-ready inference deployment.
Computer Vision Shared focus: Computer Vision.
CARE-SIU
A research framework for temporal and multimodal social-interaction understanding that combines R3D-18 video encoding, reliability-aware fusion, calibration, explainability, synthetic-to-real evaluation, and leakage-resistant experimentation for safety-critical environments.
Computer Vision Shared focus: Computer Vision.
Predictive Maintenance
Asset health, failure-risk forecasting and maintenance planning.
1 project
Technologies used
Similar use cases
Predictive & Preventive Maintenance for Generator Reliability
A full-stack analytics and ML dashboard that measures PM effectiveness, forecasts 14/30-day generator failure risk, and recommends prioritized maintenance actions.
Predictive Maintenance Shared focus: Predictive Maintenance · Generator Reliability.
Generative AI
LLM, multimodal and agentic systems for content and decisions.
9 projects
Technologies used
Similar use cases
ASHU Mentor AI Studio
ASHU Mentor AI Studio is a local-first AI mentor platform that connects resume-aware interviews, JD-based evaluation, adaptive training, evidence capture, consent-based voice generation, and digital-human lecture rendering into one complete learning workflow.
Generative AI Shared focus: Generative AI · Agentic AI · Digital Human · Voice AI.
DACR-Q
A training-free inference framework that compensates INT4 quantization error with dynamic low-rank residual correction for memory-constrained LLM deployment.
Generative AI Shared focus: LLM Inference · Quantization · Low-Rank Adaptation.
Vehicle-Scale LLMs
A memory-efficient in-vehicle LLM inference pipeline combining INT4 quantization with low-rank residual compensation for constrained edge hardware.
Generative AI Shared focus: LLM Inference · Quantization · Low-Rank Adaptation.
Autonomous Microservice Composition
An agentic MCP control plane that converts natural-language intent into executable service DAGs using registry metadata, schema retrieval, orchestration, retries, and telemetry.
Generative AI Shared focus: Agentic AI · MCP.
DeltaCert-Agent
A change-impact-aware security recertification framework for evolving tool-using LLM agents that maps configuration changes to affected assurance claims, selects scoped tests, executes risk-triggered sentinels, and escalates to broader recertification when impact cannot be bounded safely.
Generative AI Shared focus: Agentic AI.
MCP 2.0
A multi-agent platform for orchestrating tools, memory, context, knowledge, and distributed workflows over gRPC and Protocol Buffers.
Generative AI Shared focus: Agentic AI · MCP.
AegisFlow
An agentic DevSecOps cockpit for Azure Function and Python API repositories that validates CI/CD readiness, runs quality and security gates, explains failures, proposes human-approved fixes, and generates downloadable audit-ready evidence packs.
Generative AI Shared focus: Agentic AI.
MedClaim Sentinel
A local-first medical insurance claim review platform that reads multilingual claim documents, detects repeated claims and invoices, reconciles amounts, performs conservative diagnosis–medicine–bill–investigation matching, and supports evidence-grounded reviewer chat.
Generative AI Related capabilities: Multimodal AI · Conversational AI.
Execution-Aware Agentic VRP Solver and Benchmark Studio
An end-to-end execution-aware Vehicle Routing Problem platform with 8 solver backends, 18 operational scenarios, and an Agentic AI layer powered by Google Gemini 2.0 Flash.
Generative AI Shared focus: Agentic AI.
Logistics
Routing, field operations and resource allocation workflows.
1 project
Technologies used
Similar use cases
Execution-Aware Agentic VRP Solver and Benchmark Studio
An end-to-end execution-aware Vehicle Routing Problem platform with 8 solver backends, 18 operational scenarios, and an Agentic AI layer powered by Google Gemini 2.0 Flash.
Logistics Shared focus: Optimization · Vehicle Routing.
Healthcare
Clinical, insurance and medical-document intelligence.
1 project
Technologies used
Similar use cases
MedClaim Sentinel
A local-first medical insurance claim review platform that reads multilingual claim documents, detects repeated claims and invoices, reconciles amounts, performs conservative diagnosis–medicine–bill–investigation matching, and supports evidence-grounded reviewer chat.
Healthcare Related capabilities: Healthcare AI · Clinical Decision Support · Claims Intelligence · Insurance AI.
Security
Security automation, identity, trust and privacy-preserving systems.
2 projects
Technologies used
Similar use cases
LightDID-ZKP
LightDID-ZKP is a research framework introducing CAPS-ZK, a policy- and resource-aware selector for BBS and AnonCreds privacy-preserving verifiable presentations.
Security Shared focus: Decentralized Identity · Verifiable Credentials · Zero-Knowledge Proofs · Privacy-Preserving Identity.
AegisFlow
An agentic DevSecOps cockpit for Azure Function and Python API repositories that validates CI/CD readiness, runs quality and security gates, explains failures, proposes human-approved fixes, and generates downloadable audit-ready evidence packs.
Security Shared focus: Security · DevSecOps.
Research Automation
Research prototypes, benchmark workflows and reproducible studies.
2 projects
Technologies used
Similar use cases
DACR-Q
A training-free inference framework that compensates INT4 quantization error with dynamic low-rank residual correction for memory-constrained LLM deployment.
Research Automation Shared focus: LLM Inference · Quantization · Low-Rank Adaptation · Memory Efficiency.
Vehicle-Scale LLMs
A memory-efficient in-vehicle LLM inference pipeline combining INT4 quantization with low-rank residual compensation for constrained edge hardware.
Research Automation Shared focus: LLM Inference · Quantization · Low-Rank Adaptation.