Dev Bansal

Ocally Inc., Remote

• Design and ship the evaluation suite for a 19-agent research pipeline, comprising a grading engine with multiple scoring metrics, Langfuse trace integration, and regression fixtures built from real production run traces.
• Scoped the suite as a self-contained evals/ module, enabling grading to be added to a live pipeline without modifying any existing source files.
• Built a deep research agent and a business research agent that gather structured per-store intelligence automatically, replacing manual analyst lookups.
• Extract and model historical time-series sales and inventory data from POS systems (Clover, Square), covering product-level movement, basket affinity, and time-of-day, week, and month demand patterns to feed downstream trend forecasting.

Code Help, Remote

• Contributed to the launch of CodeHelp One, coordinating across engineering and marketing teams to meet product milestones.
• Managed Twitter growth and ran creator outreach campaigns securing promotional partnerships.

Built with Python, YOLOv8, OpenCV, Ultralytics

• Built a detection pipeline running YOLOv8 inference over traffic imagery across 80 COCO classes, firing severity-graded alerts (HIGH / MEDIUM / LOW) for pedestrians on road, congestion, and large vehicles.
• Encoded alert logic in a declarative rule table of watched classes, trigger thresholds, messages, and severities, allowing new rules to be added with no changes to inference code.
• Decoupled detection, alerting, and I/O into independent modules producing colour-coded terminal alerts and an annotated output image.

Built with Python, LangGraph, LangChain, Ollama, Tavily

• Engineered a stateful LangGraph workflow in which a Planner node decomposes a question into targeted search queries, a Researcher node executes them, and a Generator node synthesizes a cited report.
• Ran LLM planning and generation locally through Ollama (Llama 3) to keep prompts off third-party servers, with Tavily supplying real-time web retrieval.

Built with Python, scikit-learn, Streamlit

• Deployed a web platform that ingests an uploaded CSV or sample dataset, takes a target column, and automatically detects whether the task is classification or regression.
• Trains and compares multiple scikit-learn models in a single click, reporting metric tables, trend charts, and interactive confusion matrices.

Built with Python, PyTorch

• Reimplemented YOLOv1 (Redmon et al., 2015) directly from the paper: 24 convolutional layers plus 2 fully connected layers, 448×448 input, and a 7×7×30 output grid encoding two boxes and 20 class probabilities per cell.

Built with Python, PyTorch, torchvision

• Benchmarked three transfer-learning regimes on an ImageNet-pretrained ResNet-18: feature extraction (~5K trainable parameters), partial fine-tuning of layer3 and layer4 (~10.5M), and full fine-tuning (11.2M).
• Achieved 77.2% test accuracy on the feature-extraction baseline using differential learning rates for backbone and head, best-checkpoint selection on validation accuracy, gradient clipping, and ImageNet-normalized 224×224 inputs.