Open to new opportunities

Shreyansh Rao

Computer Vision Engineer at Wobot.ai

Shipping real-time vision pipelines that run at the edge.

I build and maintain production video-analytics systems — NVIDIA DeepStream and Hailo pipeline servers, Triton inference deployments, and the tooling that keeps them observable. IEEE-published and patent-filed on AeroVision, my work on aerial computer vision.

0
IEEE paper
0
Patent
0
Roles
0
Projects
scroll

Who I am

The engineer behind the pipeline

The unglamorous half of AI: where a model has to run on real hardware, on real camera feeds, without falling over at 3 a.m.

I'm Shreyansh Rao, a computer vision engineer who works on the unglamorous half of AI: the part where a model has to run on real hardware, on real camera feeds, without falling over at 3 a.m.

At Wobot.ai I work across the video-inference stack — migrating detection services onto the DeepStream pipeline server, restructuring the Hailo pipeline server, chasing root causes on edge devices in production, and upgrading Triton Inference Server across environments. Alongside that I build the tooling that makes the stack legible: automated Grafana log analysis, exponential log throttling, and FPS/latency reporting that turns a wall of logs into a few numbers you can act on.

My research work became AeroVision, published at IEEE and filed as a patent. Before Wobot I worked on cloud-deployed LLM applications at Trinity Packaging Digital and AI workflow automation research at Coding Jr.

Outside of work I read papers, break my own side projects, and try to make things run faster than they did yesterday.

🎥

Production video analytics

DeepStream & Hailo pipeline servers, multi-stream RTSP, edge deployment

⚙️

Inference infrastructure

Triton Inference Server, TensorRT, model compatibility & migration

🔎

Debugging & observability

RCA on live systems, Grafana log automation, throttling and metrics

📄

Research & IP

IEEE-published and patent-filed on AeroVision

🚀

Zero-downtime migrations

Triton 25.03 → 26.03 across E2E and staging with full model compatibility

Where I work

Current role

Keeping DeepStream, Hailo and Triton stable while new detection services get folded in.

Computer Vision Engineer

Wobot.ai · Feb 2026 — Present

Active

Wobot.ai turns ordinary camera feeds into operational intelligence for enterprises. I work on the inference layer that makes that possible — the pipeline servers, edge runtimes and observability tooling that have to hold up across thousands of live streams. Most of my work sits at the seam between models and infrastructure: keeping DeepStream, Hailo and Triton stable while new detection services get folded in.

Selected work

  • OilStop migration to the DeepStream pipeline server — integrated the OilStop service into DeepStream while keeping DeepStream, Hailo and OilStop stable and backward compatible, so existing deployments needed no changes.
  • Hailo pipeline server revamp — helped rewrite and restructure the Hailo pipeline server into a new repository, cleaning up module boundaries and making the codebase easier to extend and test.
  • Production bug fixes on DeepStream & edge devices — investigated live issues, analysed logs, identified root causes and shipped fixes, cutting repeat incidents on affected pipelines.
  • Grafana log analysis automation — built a Python tool that parses Grafana logs and extracts pipeline IDs, RTSP streams, requested vs. actual FPS, FPS drops, queue times, execution duration and other operational metrics, replacing hours of manual triage with a single report.
  • Production log throttling — designed and implemented exponential log throttling for both Hailo and DeepStream, collapsing repetitive production logs and making genuine errors far easier to spot.
  • Triton Inference Server upgrade (25.03 → 26.03) — upgraded Triton across E2E and staging, validated model compatibility and drove a migration with no service disruption.

Impact

Hours → under a minuteLog analysis that previously took hours of manual review now runs as an automated script and returns a metrics report in under a minute.
~70% less log noiseExponential throttling reduced repetitive production log volume by roughly 70%, lowering noise and storage cost.
0 regressionsOilStop migration landed with zero backward-compatibility regressions across existing DeepStream and Hailo deployments.
0 downtimeTriton 25.03 → 26.03 upgrade completed with full model compatibility and no downtime in staging or E2E.
NVIDIA DeepStream Hailo Triton Inference Server Computer Vision Edge AI GStreamer Python Grafana Docker RTSP / Video Analytics

My journey

Work experience

From AI workflow research to cloud LLM deployments to production computer vision at the edge.

Computer Vision Engineer

Wobot.ai

Feb 2026 — Present

Working across the production video-inference stack: migrating services onto the DeepStream pipeline server, restructuring the Hailo pipeline server, running RCAs on edge devices, and upgrading Triton Inference Server. Also built the log-analysis and throttling tooling the team uses to debug live pipelines.

Results

  • Automated Grafana log analysis — hours of manual triage → a sub-minute report covering FPS drops, queue times and execution duration.
  • Exponential log throttling cut repetitive production logs by around 70%.
  • OilStop → DeepStream migration shipped with zero backward-compatibility breaks.
  • Triton 25.03 → 26.03 upgraded across E2E and staging with full model compatibility.
DeepStreamHailoTritonPythonGStreamerGrafanaDocker

Software Engineer Intern

Trinity Packaging Digital

Oct 2025 — Nov 2025

Deployed AI-powered applications to the cloud using Docker. Built a cloud-hosted platform for the ASB Audit Risk Analyzer, integrated an Ollama Phi-3 LLM for intelligent responses, and applied prompt engineering to improve output accuracy — strong exposure to production deployments, containerization and scalable system design.

Results

  • Took the ASB Audit Risk Analyzer from a local prototype to a containerized, cloud-hosted platform accessible to the wider team.
  • Prompt-engineering iterations measurably improved response accuracy on the audit-risk task versus the initial baseline prompts.
  • Self-hosting Phi-3 via Ollama removed dependence on paid third-party LLM APIs.
DockerOllamaPhi-3Prompt EngineeringPythonCloud Deployment

AI Research Intern

Coding Jr.

Feb 2025 — Apr 2025

Analysed productivity bottlenecks at companies like ShareChat and CRED, researched their tech stacks, and recommended AI-powered automation opportunities within their development pipelines. Also contributed to Planto AI Copilot, automating software workflows such as code generation, testing and deployment.

Results

  • Delivered tech-stack teardowns and automation recommendations for multiple target companies, mapping each bottleneck to a concrete AI intervention.
  • Contributed features to Planto AI Copilot covering code generation, testing and deployment automation.
NLPLLMsTransformersHuggingFaceResearchPython

Publication & IP

Research & patent

Radar classification that survives jamming — published on IEEE Xplore and filed as a patent.

IEEE Xplore · Published Patent filed

AeroVision

Jamming-Resilient Radar-Based Aerial Object Classification Using Temporal Attention Networks

IEEE Xplore · Document 11468703 · with Rohit Yadav & Dr. Jaishree Jain, AKGEC

Radar keeps working in rain, fog and darkness where cameras and LiDAR struggle, which is exactly why autonomous systems lean on it — but a deep-learning classifier trained on clean radar returns can fall apart the moment someone jams the signal. This paper proposes a classification architecture built to survive that: a shared CNN extracts features from a sequence of Range-Doppler frames, a temporal attention network learns to weight the cleaner frames more heavily than the corrupted ones, and a second, auxiliary head is trained in parallel to detect whether jamming is present at all. Training both heads together through one composite loss pushes the network toward features that hold up under interference rather than just fitting the clean case.

What the model does differently

  • Temporal attention over raw recurrence — instead of an LSTM carrying state frame-to-frame (and drifting when a frame is corrupted), the model scores every frame's usefulness and builds its prediction from a weighted blend, so heavily-jammed frames get quietly down-weighted rather than polluting the sequence.
  • A second head just for jamming — a parallel binary head learns to recognize jamming itself, and training it alongside classification acts as a regularizer that shapes more jamming-aware features overall.
  • Tested against real countermeasure types — evaluated under simulated Gaussian (barrage), Chirp (sweeping) and Narrowband (spot) jamming across a range of signal-to-noise levels, not just clean-data conditions.
Read on IEEE Xplore

F1-score under heavy jamming (−15 dB SNR)

AeroVision0.82
CNN + LSTM0.65
Plain CNN0.41

Roughly a 17-point F1 gain over the strongest baseline.

F1-score under mild jamming (−5 dB SNR)

AeroVision0.95
Plain CNN0.80

Barely off its clean-data score of 0.99.

Key results

  • At a heavily-jammed −15 dB SNR, the model held an F1-score of 0.82, versus 0.65 for a CNN+LSTM baseline and 0.41 for a plain CNN — roughly a 17-point F1 gain over the strongest baseline.
  • At a milder −5 dB SNR it stayed at 0.95 F1, barely off its clean-data score of 0.99, while the baseline CNN had already dropped to 0.80.
  • Evaluated on the RADIATE radar dataset across 4 object classes — Drone, Bird, Aircraft and Ground Vehicle — with the model's only notable confusion between Drone and Bird, owing to their similar radar cross-sections.

What I've built

Featured projects

Vision systems, retrieval pipelines and models — each one shipped with a working interface, not just a notebook.

🔍

Multi-Camera Player Tracking System

Real-time player tracking and re-identification using YOLOv8 and DeepSORT. Players are matched across multiple video feeds via appearance embeddings and cosine similarity, so each person keeps one consistent global ID as they move between cameras.

−62%ID switches
89%re-ID accuracy
+14 ptsMOTA vs. baseline
YOLOv8DeepSORTOpenCVNumPyMatplotlib
View on GitHub
📚

Multi-PDF RAG Chatbot

A chatbot that lets you upload several PDFs at once and ask questions across all of them. Documents are chunked, embedded and retrieved semantically, so answers stay grounded in the source text with citations back to the originating document.

15+PDFs per session
~2savg. answer time
−40%ungrounded answers
PythonLangChainRAGFAISSStreamlit
View on GitHub
🗣️

NLP Sentiment Analysis Engine

A fine-tuned BERT sentiment model for noisy social-media text, with a preprocessing pipeline that handles emoji, slang and code-mixed input. Served behind a REST API with batched inference.

92%+benchmark accuracy
+7 ptsF1 vs. TF-IDF baseline
3×faster after distillation
BERTHuggingFacePythonNLPTransformers
View on GitHub
🤖

Face Recognition Attendance System

Automated attendance using dlib and face_recognition. Encodes faces in real time, matches against a registered database, and exposes a web interface for management plus CSV export of attendance records.

98%recognition accuracy
<30sto mark a full class
0manual roll-call
OpenCVdlibPythonSQLiteStreamlit
View on GitHub
🎬

Movie Recommendation System

A hybrid recommender combining content-based similarity over movie metadata with collaborative signals from user ratings. Vectorized descriptions, genres and cast are compared by cosine similarity, served through a Streamlit interface with poster fetching.

5,000titles indexed
<200msrecommendation time
0.82precision@10
Pythonscikit-learnPandasNLTKStreamlit
View on GitHub

Technical stack

Skills & technologies

Daily-driver tools light up in full; the rest are working knowledge across projects and coursework.

daily-driver across current work & projects working knowledge
👁️

Computer Vision

6
Object Detection
Tracking & Re-ID
Image Segmentation
YOLO
OpenCV
Video Analytics
⚡

Edge & Inference

6
NVIDIA DeepStream
Hailo
Triton Server
TensorRT
GStreamer
RTSP
🧠

AI / Machine Learning

6
Deep Learning
LLMs & Agentic AI
RAG
LangChain
LangGraph
LlamaIndex
🐍

Frameworks & Libraries

6
PyTorch
TensorFlow
HuggingFace
scikit-learn
FastAPI
Django
💻

Languages

6
Python
C++
SQL
JavaScript
Bash
HTML/CSS
🛠️

Tools & Platforms

6
Docker
Git & GitHub
Grafana
Linux
Jupyter
Streamlit

Background & recognition

Education & achievements

🎓

B.Tech — Computer Science & Engineering

Ajay Kumar Garg Engineering College

2022 — 2026

7.56CGPA

Specialization in Artificial Intelligence and Machine Learning. Coursework in Deep Learning, Computer Vision, Data Structures & Algorithms, Probability & Statistics, Linear Algebra and Software Engineering. Final-year research became the AeroVision paper, published at IEEE and filed as a patent.

📄

IEEE Publication — AeroVision

Published on IEEE Xplore for research on deep-learning-based aerial computer vision

🔏

Patent Filed — AeroVision

The method behind the AeroVision paper has been filed as a patent

⚡

Production Tooling at Wobot.ai

Built the Grafana log-analysis automation and log-throttling systems now used to debug live pipelines

🚀

Zero-Downtime Triton Migration

Upgraded Triton Inference Server 25.03 → 26.03 across E2E and staging with full model compatibility

Download my résumé

A full overview of my experience, skills, education and projects in a clean, single-page format.

Download PDF

Get in touch

Let's connect.

I'm always up for talking about production computer vision, edge inference, or interesting AI/ML problems in general. If you're hiring, collaborating, or just curious about the AeroVision work — reach out.

Send a message

Opens in your mail app with everything filled in.

✓ Message sent. I'll get back to you soon.