Vinay Gandhi

Data pipelines and extraction patterns for crypto-market style analytical workloads.

Market sentiment analytics — real-time data aggregation, scoring, and visualization across financial instruments.

Large-scale analytics over healthcare datasets — big-data processing and feature pipelines.

Healthcare research project — data analysis and modeling pipeline.

DETCP reference stack — MQTT → Rust edge (NATS JetStream) → gRPC → Go cloud (Kafka → TimescaleDB) for factory & fleet telemetry.

Manufacturing correlation — PLC + MES fusion and streaming ML (ingest via edge-telemetry-plane).

Manufacturing correlation engine — PLC telemetry, MES work orders, and streaming ML inference.

Conveyance & rotating equipment PM: MQTT anomalies, BM25 RAG over manuals, optional LLM diagnosis, React dashboard.

EV battery assembly vision stack: 2.5D RGB+depth fusion, VLM-style inference API, PLC halt & MES audit trail.

Dealership payment adjudication — Intake → Policy → Fraud → Adjudicator pipeline with a deterministic route_claim() gate.

Question answering over documents with retrieval-augmented generation patterns.

Agent-oriented flows for analytics workloads and structured reasoning over data.

VLM vehicle-damage assessment ETL for insurance / fleet claims — Glue → SageMaker (PaliGemma 3B) → Redshift / DynamoDB with confidence-based routing. Latency / accuracy figures are projected design targets.

AI-powered receipt OCR — extract merchant, date, totals, and line items; auto-categorize spend; dashboard analytics with CSV/Excel export. FARM-stack codebase (codename Extracta AI).

ARIMA(2,1,1) baseline upgraded to LSTM + TFT + N-BEATS ensemble — dbt feature mart, Airflow DAG, MLflow experiment tracking, FastAPI serving, Kafka streaming, and Prometheus drift monitoring.

Privacy-preserving distributed intelligence and federated learning for sensitive data silos.

Sim-to-real VLA fine-tuning profiler: WebDataset → FlashAttention-2 → FSDP. 3.3× throughput, −65% training cost per 500K-frame run.

Sim-to-real RL pipeline: PPO policy trained in Gymnasium, exported to TFLite, deployed on ESP32 for physical pendulum balance.

Ground rover with semantic scene understanding — ROS 2 Humble, LiDAR SLAM, TensorRT FP16 on Jetson Orin Nano at ≥30 FPS.

High-throughput, sub-100ms ML engine for real-time financial defense using Kappa-style streaming design.

Enterprise fleet OTA orchestration — NATS control plane, S3 data plane, Go orchestrator, and React command center.

Signature & integrity verification for OTA firmware images before they reach vehicle ECUs.

Automotive key provisioning — secure key lifecycle and distribution for vehicle ECUs.

SaaS-style expense tracker — upload receipt photos, async OCR pipeline (EasyOCR + PyTorch), MongoDB persistence, React dashboard with spend analytics and exports.

Responsive weather application — live API integration and a clean front-end build.

Lost-and-found style full-stack application with search and matching workflows.

Habit-tracking application — front-end state management and persistence patterns.

Task manager built on the FARM stack (FastAPI · React · MongoDB).

PM-authored PRD and React/TypeScript prototype for the human-in-the-loop recovery workflow when a robot's autonomous policy fails on the factory floor.

Attribute exact dollar value to human-in-the-loop robotic teleoperation data — Streamlit dashboard surfacing operator quality at per-session granularity.

B2B document-intelligence concept — receipt upload to structured expenses, categorization, and spending analytics; Dockerized FARM stack.