Hi, my name is
Sree Vidya Cheekuri.
About Me
A passionate Data Scientist and ML Engineer currently pursuing a Master's degree at the University of Houston. I build robust predictive models and AI-driven solutions, transforming complex data into actionable insights. My experience spans academic research in computer vision and NLP to developing full-stack applications.
I've always been fascinated by the stories hidden within data. For me, a spreadsheet is a mystery novel, and a neural network is the key to unlocking it. I thrive on bridging the gap between intricate theory and real-world, high-impact applications.
Beyond the code, I'm a collaborative team player who believes the best solutions come from diverse perspectives. I'm constantly exploring the latest advancements in AI and am excited to apply my skills to a challenging role where I can contribute to cutting-edge projects and continue to grow as an engineer.
My Toolkit
Programming & Databases
Data Science Libraries
AI & ML Frameworks
AI Specializations
Methods & Foundations
Tools, Platforms & Visualization
Where I've Worked
Machine Learning Research Intern @ University of Houston
Nov 2024 – Present
- Built an ensemble of Random Forest, XGBoost, and SVR models to forecast concrete strength, corrosion risk, and leaching behavior. Lab-based testing for these properties is slow and expensive; this pipeline cuts cycle time in half while maintaining strong predictive power, so researchers can iterate faster on sustainable mix designs.
- Designed a containerized safety-monitoring system powered by Vision-Language Models (Qwen 2.5-VL) and PyTorch. It scans construction imagery for OSHA-related hazards — unprotected edges, exposed wiring, missing PPE, unsafe scaffolding — using zero-shot recognition across 40+ risk categories, giving site managers a real-time compliance dashboard.
AI/ML Research Intern @ Deakin University, Australia
Mar 2023 – Jul 2023
- Implemented a tourism demand forecasting pipeline combining ARIMA (for trends and seasonality) with LSTM (for nonlinear patterns) on 5+ years of historical data. The model supports capacity planning and resource allocation for tourism operators, with versioned model artifacts for reproducibility.
- Built a Flask-backed inference service with SQL storage that serves predictions on demand. The design separates model refresh from API contracts, so newer models can be swapped in without disrupting consumers — enabling iterative deployment and continuous improvement.
Full Stack Web Developer @ Solar Secure Solutions
May 2022 – Jul 2022
- Shipped interactive web modules using HTML, CSS, JavaScript, and Python, with RESTful APIs connecting frontend and backend. The dashboards and UX improvements led to a measurable jump in user engagement across client platforms.
Things I've Built
TechSage: AI Ecosystem Intelligence
A full-stack system that pulls in AI-related news from labs, frameworks, media, and GitHub, then cleans, clusters, and ranks it. Instead of scrolling dozens of feeds, you get a unified, deduplicated view of what’s worth attention — with an optional RAG chat so you can ask questions in natural language and get answers grounded in recent articles.
- Sources span four tiers (official labs, framework vendors, tech press, GitHub trending). When the same story appears everywhere, the pipeline keeps the canonical copy and links the rest, using 384-dim embeddings and cosine similarity for semantic dedup.
- Two-stage classification: BART labels topics (LLMs, Agents, Infrastructure, etc.) and release types (Model Release, Framework Update, etc.); pgvector supports similarity search and personalized retrieval for the RAG chat.
- FastAPI + PostgreSQL backend, APScheduler for periodic ingestion (news every 5 min, GitHub every 6 hrs), and a React-style dashboard with tabs for Models, Frameworks, Agents, Research, and an Ask interface that cites sources for every answer.
6D Pose Estimation for Robotic Grasping
A deep learning pipeline that predicts 6D object pose — 3D position plus 3D orientation — from a single RGB image, mimicking warehouse-style bin picking where a robot must identify and grasp items from clutter. The system uses transfer learning: a frozen ResNet50 backbone feeds into a custom regression head that outputs translation (x,y,z) and a 6D rotation representation (more stable than quaternions for training).
- Trained on the YCB-Video dataset; the PoseNet-style model achieves sub-millimeter translation error and learns rotation in a continuous 6D space before converting to rotation matrices for evaluation. The architecture is modular and suitable for integration into robotic workflows.
- End-to-end script-based workflow: data ingestion, dataset splitting, training, evaluation, and packaging. Addressed real-world hurdles — Open3D/PyTorch conflicts on macOS, broken dataset mirrors, Git LFS for large model files — resulting in a reproducible Conda environment and clear MLOps practices.
Video Action Recognition
A PyTorch-based system that classifies human actions in video clips across 101 categories (UCF101): from sports and instruments to everyday activities. The repo supports two backbones — CNN+LSTM (2D CNN per frame, LSTM over time) and R3D-18 (3D ResNet) — with a Flask web UI for drag-and-drop prediction and a full training pipeline.
- The CNN+LSTM variant extracts spatial features per frame and models temporal dynamics with an LSTM; R3D-18 uses 3D convolutions to fuse space and time. Both accept configurable frame counts (e.g., 16 or 32) and frame sizes for tradeoffs between accuracy and speed.
- Includes data loaders, augmentation, label encoding, checkpointing, and experiment tracking. The Flask app serves real-time predictions so you can upload a video and see the predicted action instantly. Built for extensibility: adding new architectures or datasets follows a clear pattern.
Generative Storytelling with AI
An automated pipeline that turns a single text prompt into a narrated video: a Transformer-based LLM (GPT-4o) expands the prompt into a story, an image model (DALL-E 3) renders each sentence as a scene, and a TTS model produces voiceover. MoviePy stitches images, audio, and text overlays into a final video — useful for education, demos, and quick content creation.
- The LLM stage uses decoder-style attention to generate coherent multi-sentence narratives; the image stage relies on a GAN-like paradigm for text-to-image synthesis. Each sentence becomes a clip (image + audio + overlay), then clips are concatenated into the final output.
- Built with Python, FFmpeg, and ImageMagick for media handling. The pipeline is modular: swap models or adjust generation params without rewriting the assembly logic. Includes demo assets and a sample run from a "cat astronaut on a candy planet" prompt.
Multilingual Intent Classification & Slot Filling
An NLU pipeline for voice assistants and chatbots: given an utterance like "Play jazz by Miles Davis tomorrow at 6 PM," it predicts the intent (play_music) and extracts slots (MUSIC_GENRE=jazz, ARTIST_NAME=Miles Davis, DATE_TIME=tomorrow at 6 PM). Both intent and slot models are built on fine-tuned XLM-RoBERTa; slots use BIO tagging plus a CRF layer to enforce valid span transitions.
- Trained on the MASSIVE dataset (Amazon Science), which provides parallel annotations across 51 languages for 60 intents and 55 slot types. A single model handles all languages — no per-language variants. The Streamlit app runs both models in sequence: type or paste an utterance and see intent (with confidence) plus extracted slots in real time.
- Slot filling uses a token-level classification head with a CRF on top: B-ARTIST, I-ARTIST, O, etc. The CRF avoids invalid transitions (e.g., I-ARTIST before B-ARTIST). Models can be loaded from local checkpoints or Hugging Face Hub; the app works on Streamlit Cloud with lazy model loading.
Concrete Strength Prediction
Compressive strength of concrete depends on mix proportions (cement, water, aggregates, admixtures) and curing. This project predicts strength from these inputs so engineers can optimize mixes and avoid over-specification or underperforming batches. Useful for quality control and mix design iteration.
- Uses a stratified CatBoost ensemble: separate models for different strength bands to improve calibration across the full range. Domain-aware features (water-cement ratio, curing age, aggregate gradation) make predictions interpretable and align with material science.
- End-to-end pipeline from raw mix data to predictions, with feature engineering, cross-validation, and model serialization for deployment into downstream tools or APIs.
Chloride Corrosion Prediction
Steel reinforcement in concrete can corrode when chlorides (from de-icing salts or seawater) penetrate the cover. This pipeline predicts corrosion risk from mix design, exposure, and environmental factors — informing maintenance schedules, cover depth choices, and coating decisions for bridges, parking decks, and coastal structures.
- Ensemble of MLP and XGBoost: neural nets capture nonlinearity, gradient boosting adds robustness. Includes data augmentation and feature scaling for noisy real-world datasets. Sensitivity analysis helps identify which inputs drive risk most.
- Complete workflow from raw data to trained models, with clear artifact organization for reproducibility and integration into durability assessment tools.
Assistive Tech for the Visually Impaired
A MATLAB-based system that helps visually impaired users infer emotional context in conversations. Facial expressions (smiles, frowns, surprise) are often missed; this tool detects faces, classifies emotions, and delivers real-time audio feedback via text-to-speech so users can respond more naturally.
- Local Binary Patterns extract texture features from detected face regions; a neural network classifies emotions (e.g., happy, sad, neutral). The pipeline runs in near real time and pipes output to TTS for accessibility.
- Designed for practical use: robust to lighting and pose variations, with a simple interface so users can rely on it during live interactions without distraction.
Traffic Simulator for Cybersecurity Training
An NSF-supported traffic simulation platform built with SUMO and TraCI for studying cyber attacks on traffic infrastructure. Simulates vehicles, pedestrians, and traffic lights at intersections — from simple 4-way setups to the real-world Sioux Falls network — with configurable control logic.
- Includes pedestrian-actuated signals: TraCI detects waiting pedestrians and triggers green phases only when needed (after minimum vehicle green). Vehicle-actuated controller extends green when queues are present. Demonstrates how realistic signal logic can be modeled and manipulated for security research.
- Sioux Falls network: 24 nodes, 76 edges, converted from GraphML. Queue-based adaptive controller and IDM car-following. Shell scripts for quick demos; requires SUMO and SUMO_HOME. Used for cybersecurity training scenarios where attackers target signal timing or sensor data.
21-Element Cement Leaching Prediction
Predicts elemental leaching from monolithic cement materials under varying pH and time. When concrete is exposed to aggressive environments, trace elements can leach out — relevant for waste encapsulation, dams, and durability. This project forecasts cumulative release (mg/m²) for 20 elements (Al, As, Ba, Ca, Cd, Cr, Cu, Fe, K, Mg, Na, Pb, Si, Zn, etc.) from mix and exposure inputs.
- XGBoost-dominated ensemble with 25+ engineered features (interactions, polynomial terms, domain-specific metrics). Material-specific logic for problematic elements (Al, Br) with conditional ensemble selection. R² > 0.85, median absolute percentage error < 1%.
- Flask web app and REST API for predictions; deployed live on Railway. Supports multiple cement types (CEM I–V), form types (Concrete, Mortar, Sewage_Sludge), and statistical measures. Model serialization and config-driven training for reproducibility.
Job-Search Agent (Middle America AI Engineer)
An AI for Engineers assignment: an autonomous agent that finds AI Engineer roles at mid-sized “Middle America” companies (excluding FAANG+ and early-stage startups). It searches job boards, filters by company size and location, ranks by profile fit, and generates tailored resumes and cover letters for the top matches.
- Pipeline: scrape LinkedIn, Indeed, Greenhouse; blacklist 40+ big-tech firms; exclude startups via employee-count heuristic; rank by skill match, location preference, recency; tailor materials for top 3. Optional OpenAI integration for stronger generation; works fully offline with heuristic-based ranking.
- React frontend, FastAPI backend. Includes design doc, requirements, benchmark dataset (20 jobs), evaluation metrics (Precision@10, Interview Yield), ethics/bias analysis, filter-toggle experiments. Built for the AI for Engineers course with full trace logging for every filter and rank decision.
Publications
Pneumonia Detection using Machine Learning
- A CNN-based pipeline for pneumonia detection from chest X-rays, validated on labeled medical datasets with input from clinical advisors. The system targets radiology workflows where AI can flag suspicious cases for prioritization, not replace human judgment.
- Published at IEEE 5th INCET 2024 after peer review. The paper documents model design, validation protocol, and findings, adding to the literature on AI-assisted diagnostic tools in medical imaging.
What's Next?
I'm actively seeking new opportunities and collaborations where I can contribute to innovative, data-driven products.
My inbox is always open, whether you have a question or just want to say hi.
Feel free to reach out — I'm happy to chat about AI, tech, or anything in between!