Dev Sanghvi

AI engineer and co-founder of DHI

I build computer vision that runs on small GPUs next to the camera. At DHI I wrote most of our platform: 31 video analytics use cases on a single 8 GB edge GPU, a cloud dashboard, and a 12-product research program behind it. I finished my Master of Computer Science at Rice in August 2026.

DHI is my main work. I also take on a small amount of contract work: AI automation, agents, internal tools and MVPs.

ResumeGitHub LinkedIn
Rice University logoRice UniversityDHI logoDHIVMukti logoVMuktiISRO logoISROCASML 2024
Dev Sanghvi portrait

Working style

  • I think about the hardware and the deployment before the model.
  • I publish results with what they do not show.
  • I take research far enough that it runs on the box.

Best-fit roles

AI EngineerEdge AI EngineerComputer Vision EngineerInference / Deployment EngineerApplied AI / Production MLFounding EngineerContract: AI automation, agents, MVPs

Core strengths

Edge AIComputer VisionOCRReal-time Video AnalyticsDeployment OptimizationC++ / Python EngineeringTransformersPrivacy-aware MLStartup ExecutionTechnical Leadership

Building

DHI

Co-founder. DHI turns the cameras a site already has into live video analytics, running on a small local edge node. I wrote most of the platform.

Visit DHI

Research

DHI Labs

A 12-product applied research program: one repo per product, tests against synthetic ground truth, and results published with their misses.

See DHI Labs

Recently

Rice University

Finished a Master of Computer Science (AI) in August 2026, then spent the summer building DHI in Rice's Summer Venture Studio.

31

configurable video analytics use cases on 12 detector engines in the DHI platform

710 MB

memory for six live cameras on an 8 GB edge GPU, down from about 3.4 GB

10K+

IoT cameras reached by deployed analytics across large surveillance environments

2-5x

faster inference delivered while retaining more than 95% of original accuracy

98%+

multilingual detection accuracy reported for the ISRO privacy-preserving email pipeline

40%

training-time reduction in the metasurface publication presented at CASML 2024

Experience

Work across startups, production AI, and research.

Four years of video analytics in industry, research at ISRO and IIT Bombay, and now my own company.

DHI Technologies, Inc. logo

Co-Founder

Startup

DHI Technologies, Inc.

Nov 2025 - PresentHouston, TX

Building DHI, video analytics that run on a local edge node next to the cameras a site already has. I wrote most of the platform and run go-to-market.

  • Built the video analytics engine (Python, FastAPI): 31 configurable use cases, including license plates, re-identification and intrusion, on 12 detector engines.
  • Designed the multi-camera inference pipeline: six live cameras in 710 MB on an 8 GB edge GPU, down from about 3.4 GB.
  • Built the multi-tenant cloud dashboard (React, TypeScript, Cloudflare Workers, D1) and lead DHI Labs, the 12-product research program.
TensorRTFastAPIReactCloudflare WorkersEdge AIStartup Leadership
dhi-tech.com
Rice Summer Venture Studio (Liu Idea Lab) logo

Founder, DHI

Accelerator

Rice Summer Venture Studio (Liu Idea Lab)

May 2026 - Aug 2026Houston, TX

Rice's summer accelerator. DHI was selected for the cohort with a $15,000 award, and I spent the summer building it with my two co-founders.

  • Worked on DHI's product and go-to-market through the program, alongside the engineering.
Go-to-marketProductStartup Leadership
Rice University School of Engineering and Computing logo

Teaching Assistant

Academic

Rice University School of Engineering and Computing

Jan 2026 - May 2026Houston, TX

Supported course delivery and gave students structured feedback and guidance during and outside class.

  • Strengthened constructive feedback and professional communication through direct student support.
Constructive FeedbackProfessional CommunicationTeaching
VMukti Solutions logo

AI Solutions Developer

Industry

VMukti Solutions

Jul 2021 - Jan 2025Ahmedabad, India

Worked on scalable cloud VMS systems, edge AI camera software, intelligent video analytics, anomaly detection, and large-scale deployment pipelines.

  • Built production-ready C++ / ONNX inference systems, achieved 50-70% latency reduction, and delivered 60-80% model-size reduction with 2-5x faster inference.
  • Integrated GenAI video-forensics workflows with more than 95% precision and supported analytics across 10K+ IoT cameras with sub-second latency pipelines.
C++ONNXPythonVideo AnalyticsDeploymentEdge AI
vmukti.com
Adiance Technologies Pvt. Ltd. logo

Team Lead

Industry

Adiance Technologies Pvt. Ltd.

Jan 2025 - Jul 2025Ahmedabad, India

Led work on edge AI CCTV products and deployment pipelines for surveillance environments.

  • Architected and deployed Florence-2 and TimeSformer pipelines for sub-second real-time activity tracking on edge video.
  • Cut authentication latency by 40% via C++ ONNXRuntime face verification and collaborated with stakeholders in 20+ countries on deployment constraints.
Edge AIONNXRuntimeComputer VisionProduct Delivery
Indian Space Research Organisation (ISRO) logo

Machine Learning Researcher

Research

Indian Space Research Organisation (ISRO)

Jan 2025 - Apr 2025Ahmedabad, India

Built an offline-capable multilingual email-security pipeline combining OCR, transformer-based analysis, and privacy-preserving processing.

  • Worked on malicious content detection, emotional tone detection, attachment analysis, and multilingual email understanding without cloud upload.
  • Reached 98% accuracy on internal multilingual samples and about 70% fewer false positives than the rule-based baseline.
OCRBERTRoBERTaPythonText Classification
Indian Institute of Technology Bombay logo

Deep Learning Researcher

Research

Indian Institute of Technology Bombay

Jul 2024 - Aug 2024Mumbai, India

Researched intrusion detection in IoMT networks with deep learning and reinforcement learning.

  • Trained intrusion detection models on a 77 GB dataset, reaching 98% detection accuracy.
  • Experimented with Random Forest, CNN, and Kolmogorov-Arnold Networks while handling class imbalance and training / inference efficiency.
Reinforcement LearningDeep LearningCybersecurityAlgorithm Optimization
Adani University logo

Research Intern

Research

Adani University

Jul 2023 - Aug 2023Ahmedabad, India

Worked on nanotechnology, metamaterials, camouflage, and design research.

  • This internship became the foundation for the later metasurface publication presented at CASML 2024.
NanotechnologyMetamaterialsDesign Research
Adani Student Programming and Development Club logo

Vice President

Leadership

Adani Student Programming and Development Club

Mar 2023 - Aug 2024Ahmedabad, India

Scaled the largest programming club on campus through mentorship, workshops, hackathons, and repository stewardship.

  • Organized 10+ workshops and hackathons, mentored 250+ students, and helped the club reach 400+ members with demand regularly exceeding capacity.
  • Managed the club GitHub repositories and strengthened the campus coding culture.
MentoringGitHubCommunity BuildingLeadership
aspdc.tech

Project library

DHI Labs products, Rice research, and earlier projects.

Public repositories are linked. DHI Labs code is private, so those cards describe the work and the measured results.

Showing 25 of 25 projects.

DHI LabsPrivate work

Edge Scene Graphs

DHI Labs · 2026

A layer between any detector and an alerting system that turns boxes into a queryable scene graph and explainable alerts.

  • Spatial and temporal predicates, an interval-compressed graph, and a rule engine whose alerts point back to the evidence.
  • 2.4 ms p50 per frame on-device; learned relation head R@20 0.500 against 0.220 for a frequency baseline on Visual Genome.
Scene GraphsRule EngineVisual GenomeTensorRTPython
DHI LabsPrivate work

Cross-Camera Identity Memory

DHI Labs · 2026

Links people across cameras and answers questions like where someone went, from any tracker's event stream.

  • Prefers leaving a link out to making a wrong one, and keeps a bounded memory that survives restarts.
  • Zero wrong links across 7 cameras and 313 people on WILDTRACK; ships with a Frigate adapter.
Multi-cameraRe-identificationSQLiteWILDTRACKFrigate
DHI LabsPrivate work

Continual Open-Vocabulary Detection

DHI Labs · 2026

Add a new class on-device from a text prompt or a handful of image crops, with no gradient updates.

  • A gate replays the site's evaluation set before a new class goes live, and classes can be rolled back.
  • On COCO val2017, 4 of 6 added classes went live with zero forgetting of existing classes.
Open VocabularyCLIPContinual LearningCOCO
DHI LabsPrivate work

Occlusion-Aware Counting

DHI Labs · 2026

People counts that correct for who the detector cannot see, with calibrated intervals instead of a single number.

  • Visibility estimation, detectability calibration and conformal 90% count intervals.
  • MAE 2.46 against 3.73 for naive counting in heavy crowds on synthetic ground truth, then tested on CrowdHuman.
CountingConformal PredictionCrowdHumanPython
DHI LabsPrivate work

Predictive Alerting with a Falsification Ledger

DHI Labs · 2026

Forecasts incidents a few seconds ahead, explains each forecast by replaying it with a cause removed, and scores every forecast against what happened.

  • A 210-scenario battery run through the production code, reported per alert type, including the weak one.
  • 8% false alarms across the no-incident scenarios.
ForecastingCounterfactualsEvaluationPython
DHI LabsPrivate work

Fixed-Camera 3D

DHI Labs · 2026

Calibrates a fixed camera from people walking past, with no calibration target, then reports distance, speed, height and density.

  • 0.16 m ground-position RMSE for a 5 m, 30 degree mount on synthetic ground truth.
Camera Calibration3D GeometryPython
DHI LabsPrivate work

Thermal Perception

DHI Labs · 2026

A thermal perception engine that needs no training, plus the data and pretraining tooling for thermal models.

  • Masked-autoencoder pretraining pilot on real LWIR images (LLVIP), reported as training mechanics rather than accuracy.
ThermalMAE PretrainingViTLLVIP
DHI LabsPrivate work

Prompt2Model

DHI Labs · 2026

Turns a plain-language request into a trained, compressed model ready for an edge device.

  • Planner, training, ONNX export and evaluation in one pipeline.
  • Refuses to ship a compressed model that falls below its accuracy floor and keeps the uncompressed one instead.
LLM PlannerModel CompressionONNX
ResearchPublic repo

Collision-Risk Forecasting with Spatiotemporal Graphs

Rice University (team of 3) · Spring 2026

Forecasts calibrated collision probability and time to event for autonomous driving, instead of raw trajectories.

  • Led the graph engine and models: kNN, radius and time-to-event graph builders, ST-GNN (GAT + GRU) and graph transformer encoders.
  • Multi-head decoder for risk and time to event, plus attribution tools to explain predictions.
GNNsPyTorchHydraAutonomous DrivingCalibration
View code
ResearchPublic repo

MemFaith: Context Faithfulness in Long-Context LLMs

Rice University · Spring 2026

Measures how much each chunk of context drives a model's answer, by leaving chunks out one at a time.

  • Deterministic chunk ablation on FEVER-style and HotpotQA-style examples, with a Hugging Face Transformers backend.
LLM EvaluationFEVERHotpotQATransformers
View code
Data & PrivacyPublic repo

DP-accurate DAU/MAU Counter Under Deletions

Rice University · Sep 2025 - Dec 2025

Deletion-aware distinct counting for privacy-sensitive analytics and compliance-heavy measurement.

  • Designed a deletion-aware pipeline using KMV sketches, tombstone propagation, and RDP accounting for accurate DAU/MAU measurement after deletion events.
  • The work centered on correctness, auditability, and practical GDPR-compliant analytics.
PythonKMV SketchesRDPPrivacy AnalyticsSQLBackend Engineering
View code
ResearchPublic repo

Editability and Faithfulness Metric for FEVER

Rice University · Sep 2025 - Dec 2025

LLM evaluation work focused on editability-faithfulness tradeoffs on FEVER-style claims.

  • Benchmarked GPT-2 and ROME edits across valid examples, comparing target and control flips to quantify whether edits remained localized and reliable.
  • The evaluation was built to stay measurable, reproducible, and easy to interpret.
GPT-2ROMELLM EvaluationNLPTransformersExperimental Design
View code
Edge AINo public repo

Activity Identification and Triggering System

Ahmedabad Research Project · Jan 2025 - Apr 2025

Real-time video activity detection and event-triggering platform for surveillance-style streams.

  • Built an end-to-end pipeline with YOLOv11, TimeSformer, FastAPI, React, and WebSockets for live inference, configurable thresholds, and dashboard monitoring.
  • Reached about 95% validation accuracy and was shaped for operational low-latency use cases.
YOLOv11TimeSformerFastAPIReactWebSocketsEdge AI
Edge AIPrivate work

Face Recognition Pipeline

VMukti / Adiance · 2024 - 2025

Face verification pipeline built for practical edge deployment and runtime stability.

  • Worked with RetinaFace detection, ViT embeddings, ONNX / ONNXRuntime inference, CVLFace / AdaFace migration, and FRVT-aligned verification goals.
  • Focused on low FMR / low FNMR targets, C++ deployment constraints, and deployable runtime behavior.
RetinaFaceViT EmbeddingsONNXONNXRuntimeC++Computer Vision
Edge AIPrivate work

Live Video Analytics with Florence-2 and SAM

VMukti / Adiance · 2025

Interactive live-stream intelligence with multimodal video understanding components.

  • Implemented real-time pipelines for activity tracking, person tracking, object tracking, and live feed summarization.
  • Built around Florence-2 and SAM-style components in a deployable analytics flow for CCTV and edge environments.
Florence-2SAMVideo AnalyticsTrackingMultimodal AIPython
ResearchPublic repo

Vision Transformer for Accelerated Design of Very Low-frequency Metasurface Absorber

Adani University · Sep 2023 - Sep 2024

Research project combining transformers with materials / physics design optimization.

  • Used Vision Transformer style modeling for image regression on metasurface absorber data and improved design exploration over CNN-like baselines.
  • Reduced training time by about 40% and later presented the work at CASML 2024, IISc Bangalore.
Vision TransformersImage RegressionResearchPythonCNNPublication
View code
ProductPublic repo

GUI-Based Machine Learning Model for Interactive Learning

Adani University · May 2023 - Aug 2024

Interactive educational ML platform aimed at non-expert users.

  • Focused on GUI development, accessibility, and making machine learning easier to explore in a classroom context.
  • It also reflects usability work and product-minded development for educational software.
Machine LearningGUI DevelopmentEducational TechPythonAccessibility
View code
ResearchPublic repo

Deep Learning-based IoT Security Model

Indian Institute of Technology Bombay · Jul 2024

Intrusion-detection work in an IoMT / IoT security setting.

  • Optimized Random Forest and CNN approaches while also exploring hyperparameter tuning, ensemble ideas, and broader efficiency concerns.
  • Part of the larger IIT Bombay research thread around cybersecurity, anomaly detection, and real-world security data.
CybersecurityIoMTRandom ForestCNNDeep Learning
View code
Edge AIPublic repo

Text Detector for 16-Segment Displays

VMukti Solutions · Jan 2024 - Apr 2024

Industrial OCR project for noisy 16-segment display text extraction.

  • Used OpenCV, EasyOCR, and Flask to support real-time text detection from images and videos.
  • It involved OCR, backend integration, and practical computer vision under messy visual conditions.
OCROpenCVEasyOCRFlaskComputer Vision
View code
Data & PrivacyPublic repo

OCR-Based Emotion Detection

ISRO · Jan 2025 - Apr 2025

Privacy-aware email understanding project combining OCR with language embeddings.

  • Used EasyOCR and RoBERTa embeddings to classify emotional tone in email content.
  • It also involved privacy-sensitive multimodal processing in a government research setting.
OCRRoBERTaText ClassificationPythonEmail Intelligence
View code
ResearchNo public repo

Multimodal Song Emotion Classifier

Personal Project · Sep 2024

Multimodal emotion-classification pipeline combining audio and lyric signals.

  • Used Librosa audio features plus SentenceTransformer lyric embeddings with TensorFlow.
  • It combined feature fusion with multimodal emotion prediction.
TensorFlowLibrosaSentenceTransformerMultimodal AIAudio
ProductPublic repo

Alumni Management System

Adani University · 2023

University-focused platform for alumni records, networking, and engagement.

  • Built around backend/system design plus UX thinking for campus-scale software.
  • It also reflects full-stack product delivery outside core AI work.
PythonDjangoPostgreSQLWeb AppUX
View code
ProductPublic repo

Hospital Management System

Adani University · 2023

Healthcare operations project focused on adoption and usability.

  • Included chatbot-style assistance to improve user interaction with HMIS workflows.
  • The project focused on making day-to-day HMIS workflows easier to use.
ReactNode.jsChatbotHealthcareProduct Thinking
View code
ProductPublic repo

Stock.Pi

Adani University · 2023

Market sentiment analysis and chatbot system combining scraping and real-time processing.

  • Brought together scraping, sentiment analysis, and market-oriented data processing in one workflow.
  • It also highlights finance-flavored data work, chat-based interfaces, and end-to-end delivery.
Web ScrapingSentiment AnalysisNLPPythonFinance
View code
ProductPublic repo

Amazon Scrapper

Adani University · 2022

Scraping-oriented backend project that reinforced practical data collection fundamentals.

  • Built a crawler and dashboard flow for Amazon product data collection.
  • It helped build practical scraping and backend fundamentals.
Node.jsPuppeteerWeb ScrapingReactMongoDB
View code

Publication and awards

Publication, whitepapers, and awards.

The published paper, the DHI whitepapers, the accelerator and competitions.

2024CASML 2024, IISc Bangalore

Vision Transformer for Accelerated Design of Very Low-frequency Metasurface Absorber

Applied a Vision Transformer architecture to optimize metasurface absorber design and reduced training time by 40%.

40%
training-time reduction
Vision TransformersResearchMetasurface DesignImage Regression
View related repo
May 2026Rice University, Liu Idea Lab

Summer Venture Studio cohort, $15,000 award

DHI was selected for Rice's summer accelerator, which ran from May 18 to August 7, 2026.

2026Texas A&M Mays Business School

Semifinalist (top 27), AI Venture Velocity Challenge

DHI reached the semifinal round and submitted its stage-two experiment logs.

Jul 2026DHI Labs

Two DHI whitepapers

Privacy-native edge video analytics, and a hybrid edge-cloud VMS architecture.

Dec 2022IIT Bombay

1st Place, E-Yantra Robotics Workshop

Designed and programmed robotic systems / robotic arms, demonstrating robotics, embedded problem-solving, and teamwork.

Nov 2022IQM Corporations

Winner, AU-IQM Hackathon

Built an AI / ML solution for data classification and placed 1st out of 50+ teams.

Technical stack

Tools, frameworks, and systems I keep reaching for.

Mostly vision, edge inference and the web and cloud pieces around them.

Machine Learning and Modeling

Training, evaluating and calibrating models, from vision to language.

PyTorchTensorFlow / Kerasscikit-learnHugging Face TransformersVision TransformersBERT / RoBERTaLarge Language ModelsGraph Neural NetworksReinforcement LearningConformal PredictionDifferential PrivacyAnomaly DetectionMultimodal AIModel Evaluation

Computer Vision and Video

Detection, tracking and understanding on live camera streams.

Real-time Video AnalyticsObject Detection (YOLO)Object and Person TrackingRe-identificationFace Recognition (RetinaFace)Scene GraphsFlorence-2TimeSformerSAMOpenCVOCR (Tesseract, EasyOCR)

Edge Inference and Deployment

Getting models to run fast on small hardware and keep running.

TensorRTONNX / ONNX RuntimeQAT / PTQ QuantizationNVIDIA JetsonGStreamerMediaMTXRTSP / ONVIFDockerLinuxCloud VMS

Backend, Web and Cloud

The APIs, dashboards and cloud pieces around the models.

FastAPIFlaskNode.jsReactNext.jsTailwind CSSCloudflare WorkersCloudflare D1REST APIsPostgreSQLSQLiteMongoDBWeb Scraping

Languages and Tools

What I write the code in, and what I analyze data with.

PythonC++CTypeScriptJavaScriptSQLGoJavaMATLABGitNumPyPandasDask

Research Areas

Topics I have published on or researched.

Edge AIPrivacy-preserving AnalyticsLLM EvaluationIoMT SecurityMetamaterialsNanotechnology

Education

Education

A Master's at Rice, finished in August 2026, and a B.Eng. in AI-ML from Adani University.

Rice University logo

Rice University

Master of Computer Science (AI focus)

Aug 2025 - Aug 2026Houston, TX
  • GPA: 3.69/4.0
  • Coursework: Deep Learning, Deep Learning for Vision and Language, NLP, Machine Learning with Graphs, Probabilistic Algorithms, Parallel Computing
  • Teaching Assistant, Spring 2026
Adani University logo

Adani University

B.Eng. in Computer Science and Engineering (AI-ML)

Sep 2021 - Apr 2025Ahmedabad, India
  • CPI: 9.34/10.00
  • SPI: 10.00/10.00 in the 8th semester

Earlier academic background

  • Earlier academic background includes NIOS high school in Physics, Chemistry, and Math.
  • 10th board background: Shreyas Foundation, 96.47 PR / A2 grade.

Professional narrative

AI engineer and co-founder of DHI, with a Master of Computer Science from Rice (2026). Four years of production video analytics before that, plus research at ISRO and IIT Bombay.

Contact

Get in touch

Email is the fastest way to reach me.

What I optimize for

  • AI that still works once latency, privacy and small hardware become real constraints.
  • Research taken far enough that people can use it.
  • Numbers that come with their limits.

Availability

DHI comes first. I take on a small number of part-time contract projects in AI automation, agents, internal tools and MVPs, outside DHI's camera and safety space.