32.78° N, 96.80° W · Dallas, Texas

Srijeet Qazi

I build localization systems that work when GPS doesn't.

I am a Ph.D. researcher in Electrical Engineering at The University of Texas at Dallas. My advisor is Prof. Mohammad Saquib. My models estimate where a person or a vehicle is from very little information. The inputs are how long someone traveled, which way they went, and what they saw.

See how it works Read the I-NAV paper Résumé
  • Two first-author publications
  • Patent filed on I-NAV
  • NSF I-Corps National, Spring 2027
Portrait of Srijeet Qazi standing in front of blue spruce trees

Inverse navigation

Start with the clues.
End with a location.

Normal navigation starts from a known location and finds a route. I-NAV works backward. It takes what a person remembers about a trip. It then narrows the map to the places they could be. Add clues below and watch the search area get smaller.

  • Start
  • Likely area

Clues applied

  1. Travel timeAbout 14 minutes
  2. DirectionMostly northeast
  3. LandmarkEnded near a lake
Likely area100%

No clues yet, so every reachable intersection is a candidate.

This is a simple illustration. It runs in your browser. The streets, the trip, and the percentages are made up, and they are not results from the paper. Likely area is the share of reachable intersections with a probability of at least 10% of the best one.

Coverage

Research

How precisely can you find someone from incomplete information?

This question runs through my work in the Wireless Communication Research Laboratory at UT Dallas. The uses are places where GPS is missing or weak. Examples are rural areas, thick forest, disaster zones, and search and rescue.

Published · Array, 2026

I-NAV for vehicles

This model estimates where a car went. It uses three inputs. The first is the set of places the car could reach. The second is how long it drove. The third is the direction it took.

A direction-limited Dijkstra search lists the paths that fit the time budget. A Gaussian likelihood model handles error in the report. I tuned seven parameters with Bayesian optimization. The tests used real GeoLife trajectories on OpenStreetMap road networks.

Intersection over union
79.5%
Likelihood precision
0.702
True destination kept in region
94 of 94 trials
Baselines outperformed
Routing, KDE, particle filter
  • Bayesian inference
  • Bayesian optimization
  • OSMnx
  • GeoPandas
  1. 1

    Map and trips

    Roads plus recorded drives.

  2. 2

    Travel time

    Cost per road segment.

  3. 3

    Reachable area

    Where the car could arrive.

  4. 4

    Likelihood

    Time and direction weighted.

  5. 5

    Ranked map

    Candidate locations in order.

Map and trip data go in. A ranked likelihood map comes out.
  • A 19 minute test trip. The left map uses time alone. The right map adds a southwest bearing. The green dot is the start and the red dot is the true end.

Manuscript in preparation · 2026

I-NAV on foot

This model estimates where a person on foot went. A travel-time model sets the area the person could reach. A graph neural network then ranks the locations inside that area. The network learns how people choose paths at junctions. Slope, land cover, and path type change the cost of each step.

I trained it on 2,943 recorded walks and 164,605 junction decisions. The evaluation uses 758 queries from 657 walks. It includes community walks collected in Uganda by the Bwindi Community Hospital team.

Recorded location inside reachable area
750 of 758 (98.9%)
Mean search area, learned ranking
1.13 km²
Mean search area, travel time only
4.90 km²
Area reduction
76.9%
Median search area
0.21 km²
Uganda walks inside reachable area
324 of 329
  • Graph neural networks
  • PyTorch
  • Overture maps
  • Copernicus GLO-30

  1. Training

    1. Recorded walks

      2,943 routes with terrain.

    2. Learned movement

      Five graph networks.

  2. Prediction

    1. Sparse query

      Last point and elapsed time.

    2. Reachable area

      Travel time plus pace margin.

    3. Ranked cells

      Movement propagated in time.

  3. Result

    1. Probability map

      Ranked cells in the area.

    2. Evaluation

      Scored after the prediction.

Recorded walks train the choice model. A reported bearing, when there is one, sharpens the ranking.
  • A recorded 84 minute walk. Darker cells hold the top probability mass. The star is the last known point and the red cross is the recorded end.
  • Walking speed predicted from age, sex, height, and weight, tested on 158 independent participants. Recent measured pace works better than demographics.

In progress · preliminary results

Finding a route from a description

A person describes a route in plain words. They mention slopes, turns, plants, and water. An HMM and XGBoost model compares that description against terrain, land cover, and map data. It then ranks the routes that fit.

I am adding vision-language models to pull observations out of text and images. LiDAR and elevation data add terrain detail.

Candidate routes searched
3.3 million
Exact route ranked first
37.8%
Median endpoint error
34 m
Endpoint within 90 m
About 71%
  • LiDAR
  • DEM
  • Sentinel-2
  • Hydrography
  • VLMs
  1. 1

    Prior

    Last point, time, heading.

  2. 2

    Reachable region

    Places within reach.

  3. 3

    Description match

    Report against a terrain index.

    • Terrain description
    • Slope, cover, water
  4. 4

    Location zone

    Calibrated area.

  5. 5

    Route families

    Ranked routes.

The workflow. The field input is a written terrain description. No photograph upload or image processing is needed.
  • Observed terrain in four directions, above the same views rebuilt from map data. The model compares the two to place the walker.
  • Two recorded hikes. Red is the true route. Blue shows the ranked search region that holds it.

Active · NIH-funded pilot

COMPASS

Connected One Health Mobile Platform for Advanced Surveillance Systems. COMPASS is a disease surveillance project in a rural field setting. I run the data systems and the machine learning work.

I set up and maintain the Linux, PostgreSQL, and DHIS2 server that field teams use to collect data. The models combine movement, environment, animal health, and clinical data. I am testing transformers, graph neural networks, and federated learning.

Collaborators
UC Berkeley, UC Davis, Gorilla Doctors, Bwindi Community Hospital
Infrastructure
Linux, PostgreSQL, DHIS2, ETL
My role
Data systems and AI/ML
  • The DHIS2 portal that field teams use. The server behind it is one of the systems I maintain.
  • Presenting the poster on inverse navigation for disease prevention work.

Publications

Peer-reviewed work.

Array · Elsevier · 2026

I-NAV: Inverse Navigation for questionnaire-based geolocation in GPS-denied environments

S. Qazi, J. Alejandro, R. U. Murshed, T. S. Evans, J. Lippert, M. Saquib

Vol. 31, Article 100978 · doi:10.1016/j.array.2026.100978

Read on ScienceDirect

From research to practice

  • Patent filed

    I co-developed the I-NAV inverse navigation technology with my advisor. The patent application is filed.

  • NSF I-Corps Southwest

    I completed the regional program as Entrepreneurial Lead. The work was customer discovery for GPS-free navigation.

  • NSF I-Corps National

    The team was accepted to the national cohort. It begins in Spring 2027.

The NSF I-Corps Southwest team spotlight for I-NAV.

Projects

Hardware, edge AI, and embedded systems.

Personal project · 2026

Edge AI home assistant

This assistant runs on a Raspberry Pi 5. A Qwen 2B language model runs on the device itself. It answers questions about its own sensor and event logs. It recognizes faces and flags people it does not know. It reads hand gestures. It sends an alert when temperature or humidity cross a limit. Logs go to AWS for later review.

  • Raspberry Pi 5
  • Qwen 2B
  • OpenCV
  • RAG
The Raspberry Pi and camera on the bench.

  • Edge first

    • Runs without the internet
    • Data stays in the house

  • Motion trigger

    • Frame differencing
    • Detector as second stage

  • Face recognition

    • Embedding baseline
    • Unknown person alert

  • Gestures

    • Contour and skin
    • Landmark classifier

  • Decisions

    • Rule engine
    • Anomaly detection

  • Local assistant

    • Qwen 2B on device
    • RAG over logs
How the assistant splits into parts.

Published at IWCMC · 2023 to 2024

Encrypted sensor link for utilities

Two TI CC1352P1 boards send temperature data. One board encrypts each reading with a matrix cipher. The other recovers it at 115,200 bps. The design targets underground utility monitoring during storms like Winter Storm Uri.

Semifinalist, UT Dallas Undergraduate Research Scholars Award, top 20 of 200+

  • C
  • TI CC1352P1
  • LoRaWAN
The two boards during testing.
  1. 1

    Manhole sensor

    Temperature by the pipeline.

  2. 2

    Encrypt on the node

    Matrix cipher on the CC1352P1.

  3. 3

    5G and SDN

    Routed without re-ciphering.

  4. 4

    Monitoring centre

    Alerts and analysis.

From the sensor in the manhole to the monitoring centre.

Senior design, sponsored by Toyota · 2023 to 2024

Rogue ECU detection on CAN bus

This system finds unauthorized electronic control units on a vehicle network. I worked on the detection logic and on the hardware and software integration.

  • C++
  • Verilog
  • Arduino
  • FPGA
Sponsor
Toyota
Team
Six students
My part
Detection logic and integration
Hardware
Arduino and FPGA

About

Research that has to work outside the lab.

I am a Ph.D. student at UT Dallas. I earned my B.S. in Electrical Engineering here, cum laude. My work combines probability, machine learning, and map data.

Most of it has to run in the field. One example is a server that supports clinics in rural Uganda. Another is a sensor system on a Raspberry Pi.

I have also taught engineering math to sections of more than 50 students.

Education

  • Ph.D., Electrical EngineeringThe University of Texas at DallasExpected May 2028
  • M.S., Electrical EngineeringThe University of Texas at DallasExpected Dec 2026
  • B.S., Electrical Engineering, cum laudeThe University of Texas at DallasMay 2024

Teaching

  • Teaching Assistant, ECEI taught one section of Advanced Engineering Mathematics on my own, for more than 50 students. I also supported Wireless Communications and Probability and Statistics.2024–2026

Awards

  • Undergraduate Research Scholars AwardSemifinalist, top 20 of more than 200 entries.UT Dallas
  • Two first-author papers and a filed patentArray 2026 and IWCMC 2024.

Toolkit

Languages

  • Python
  • C/C++
  • MATLAB
  • SQL
  • Verilog

Machine learning

  • PyTorch
  • TensorFlow
  • GNNs
  • Transformers
  • VLMs
  • XGBoost
  • HMMs

Probabilistic methods

  • Bayesian inference
  • Bayesian optimization
  • Particle filters
  • Conformal prediction

Geospatial

  • OpenStreetMap
  • OSMnx
  • GeoPandas
  • LiDAR and DEM
  • Sentinel-2

Data and infrastructure

  • Linux
  • Docker
  • PostgreSQL
  • DHIS2
  • AWS EC2
  • TACC HPC
  • Git

Embedded and IoT

  • Raspberry Pi
  • Arduino
  • TI CC1352P1
  • LoRaWAN
  • FPGA
  • CAN bus

Outside the lab

Extracurricular and athletics.

I play table tennis for UT Dallas and led the club for a year. I also play in campus leagues and local tournaments.

  • President, UT Dallas Table Tennis Club About 50 members. I managed a budget of more than $12,000 for travel, tournament entry, and equipment. I also set up joint events with other student groups.
  • UT Dallas competitive table tennis team I competed at the NCTTA National Collegiate Championships.
  • Intramural pickleball champion UT Dallas University Recreation.
  • The UT Dallas team at the NCTTA National Collegiate Championships.
  • With teammates at the NCTTA championships in Rockford, Illinois.
  • Intramural pickleball champion at UT Dallas.
  • Medal ceremony at a local table tennis tournament.

Contact

Let's work on something together.

I am open to research internships and to collaborations. My areas are localization, geospatial AI, autonomy, and edge machine learning.

Email me Résumé (PDF) LinkedIn

srijeet.qazi@utdallas.edu