Felix Mathew Software Engineer · ML Research
Let’s talk

Austin, Texas · Edge video & ML research

Less video sent,
more of the animal kept.

I’m Felix Mathew, a software engineer with an M.S. in Computer Science from Texas State University. My thesis research builds ROI-aware wildlife video compression for edge devices, and its extension was accepted as a short paper at IEEE ICTAI 2026.

Explore the research View résumé
97.40% Reduction in transmitted archive size
93.28% Fewer detector calls than per-frame detection
M.S. Computer Science, Texas State University · GPA 3.56/4
3 Credentials: 2 AWS certifications + 1 CodePath certificate

01 · Research & projects

Compression that respects the subject.

Wildlife cameras work with limited bandwidth, storage, compute, and power, while animals often fill only a small part of the frame. This work reduces what has to be sent without losing quality around the animal.

01 / Master’s thesis Defended April 3, 2026

Texas State University · Advisor: Dr. Vangelis Metsis

Neural ROI-aware video compression for wildlife monitoring

A research prototype that detects animals on selected frames, tracks them with optical flow, removes redundant frames, and compresses the animal region and background at different quality levels. The receiving server reconstructs the video with AMT frame interpolation, compositing, and temporal stabilization. The architecture was also evaluated on NVIDIA Jetson Orin Nano hardware.

  • Object detection
  • Optical flow
  • AMT interpolation
  • Jetson Orin Nano
Archive size reduction
97.40%
Animal-region PSNR
36.12 dB
Animal-region MS-SSIM
0.9758
Fewer detector calls
93.28%

Measured across 20 held-out wildlife clips with the selected thesis configuration. Detector-call reduction is compared with running detection on every frame.

  1. 01DetectSelected frames
  2. 02TrackOptical flow
  3. 03CompressROI + background
  4. 04ReconstructAMT + compositing
  5. 05StabilizeTemporal smoothing
02 / Research software Open source

Deterministic INT16 runtime for DCVC-RT

Reproducible neural video codec bitstreams

An INT16-oriented encoding and decoding runtime built around Microsoft’s DCVC/DCVC-RT model family, using PyTorch with CUDA/C++ integration, rANS entropy coding, and FFmpeg-based preprocessing and reconstruction. It includes SHA-256 hash comparisons, byte-level bitstream validation, runtime metadata, automated tests, and setup tooling.

Bitstream size
4.82 MB
Mean RGB PSNR
28.90 dB
Tests passed
35

2,593-frame, 1280×720, 30 FPS video; MS-SSIM 0.9336. A repeated 32-frame test produced byte-identical bitstreams in the same environment. Cross-device reproducibility needs further testing, and the runtime requires an NVIDIA CUDA GPU.

  • PyTorch
  • CUDA/C++
  • rANS
  • FFmpeg
Source code
03 / Publication Co-author

IEEE ICTAI 2026 · Short paper (accepted)

FaunaCodec: ROI-aware video compression and reconstruction for wildlife monitoring

An extension of the thesis research. For the conference work I implemented the deterministic DCVC-INT16 runtime, which addresses inconsistencies between encoding and decoding on different GPU architectures. A collaborating PhD student contributed the generative upscaling component.

  • ROI compression
  • DCVC-INT16
  • Determinism

02 · Experience

Teaching, research, and production software.

Graduate research and teaching at Texas State University, preceded by backend and full-stack engineering at NeoSOFT Technologies.

Jan 2025 — Jun 2026 Graduate Teaching AssistantTexas State University · San Marcos, TX · Promoted from Graduate Instructional Assistant

Led laboratory sessions and supported instruction for CS 1428, an introductory programming course.

  • Led CS 1428 lab sessions covering C++ programming, pointers, dynamic memory, object-oriented programming, debugging, and introductory data structures.
  • Supported grading coordination and student instruction for a course serving more than 100 students.
Thesis defended Apr 2026 Master’s Thesis ResearchTexas State University · Advisor: Dr. Vangelis Metsis · Completed May 2026

Research on ROI-aware wildlife video compression under edge-computing constraints.

  • Developed and evaluated the compression and reconstruction pipeline: detection on selected frames, optical-flow tracking, ROI and background compression, and AMT-based reconstruction.
  • Implemented the deterministic DCVC-INT16 runtime used in the conference extension.
  • Co-authored “FaunaCodec: ROI-Aware Video Compression and Reconstruction for Wildlife Monitoring,” accepted as a short paper at IEEE ICTAI 2026.
Aug 2021 — Jan 2023 Associate Software EngineerNeoSOFT Technologies · Mumbai, India

Developed backend and full-stack components for production applications.

  • Built production components using AWS, relational databases, REST APIs, and event-driven architectures.
  • Built AWS Lambda-based services and tenant-isolated workflows with MySQL.
  • Implemented idempotent processing and developed transaction-safe data recovery procedures.

03 · Open source

Fixes for AI agent and LLM infrastructure.

Bug fixes and regression tests contributed through the CodePath AI301 program. Each card states where the pull request stands.

Error handling Merged · July 2, 2026

Prebid SalesAgent

Fixed exception mapping so authentication failures, timeouts, and connection failures keep their proper error classifications instead of becoming generic adapter errors.

  • Consolidated the mapping into a shared helper used by the signals and creative agent registries.
  • Added regression tests; the contribution was approved and merged.

Pull request #1483

Gemini tool-call fallback Submitted · Not merged

LiteLLM

A proposed fix for incompatible tool-call thought signatures when the router falls back between Vertex AI and Google AI Studio Gemini endpoints.

  • Removes foreign signatures while keeping assistant tool-call IDs and tool-response IDs consistent.
  • Added unit and regression tests and addressed automated review feedback; validated at the unit-test level.

Pull request #33471

04 · Credentials

Foundation and continued practice.

Master’s GPA improved from 3.33 to 3.56, with 4.00 term GPAs in Summer and Fall 2025 and A grades in Parallel Processing, Survey of Software Engineering, Software Quality, Advanced Software Engineering Project, and Formal Methods in Software Engineering.

Education

Master of Science in Computer Science

Texas State University · San Marcos, Texas

May 2026 · GPA 3.56/4

B.Tech. in Computer Science and Engineering

Dr. Babasaheb Ambedkar Technological University

2021 · GPA 8.56/10

05 · Contact

Let’s build something that has to work.

I’m interested in software and machine-learning roles where reliability, reproducibility, and careful evaluation matter.

felix-mathew@outlook.com
Résumé LinkedIn GitHub