Mitanshu Goel is an AI and robotics systems engineer based in Delhi, India. He builds edge-deployed inference pipelines, robot perception systems, and embedded AI hardware, with expertise in ROS2, YOLOv8, PyTorch, SDXL, llama.cpp, and ESP32 firmware. He is currently seeking research engineering and robotics software roles.
Mitanshu Goel
AI & robotics systems — from embedded firmware to deployed inference pipelines, built under hardware constraints.
My work sits at the boundary between physical systems and learned models — building AI pipelines where compute budgets, latency ceilings, and hardware interfaces are non-negotiable constraints, not afterthoughts. I approach perception and inference problems from the deployment end first: what runs on a Raspberry Pi, what fits in 16GB VRAM, what survives a sensor dropout.
Current technical focus: multi-modal robot perception, edge-inference optimization, and the system architecture decisions that determine whether a model benchmarks well or operates reliably.
Final-year ECE student at MAIT Delhi (graduating 2026), building across the full stack from microcontroller firmware to deployed ML inference — with a particular obsession for systems that remain reliable when the hardware misbehaves.
- Engineered a real-time voice pipeline using NVIDIA NeMo for speech-to-text with custom wake-word detection integrated into a physical robot system — latency requirements drove all architecture decisions.
- Built a polling-based interface layer between robot hardware and an AI agent for low-latency query processing and feedback; designed for fault tolerance under intermittent hardware responses.
- Trained and deployed custom YOLOv8 models for three distinct production tasks: human tracking, package classification, and gesture-based control — each with separate training regimes and inference pipelines.
- Developed a hardware telemetry workstation on ESP32/Raspberry Pi capturing environmental sensor data for predictive analytics, bridging embedded firmware with Python processing layers.
- Configured a 6-DOF robotic arm in ROS/Gazebo, debugging URDF kinematic configurations and resolving simulation-to-real discrepancies blocking stable trajectory execution.
- Integrated MoveIt for inverse kinematics and collision-aware trajectory planning using C++; achieved 50% reduction in execution time through shortest-path algorithm selection and parameter tuning.
- Minor: Artificial Intelligence & Machine Learning
- Key coursework:
Projects
Darwin Studio
Treats SDXL latent tensors as genetic material — mutation and crossover applied between generations. Custom SLERP for geometrically consistent latent interpolation.
Memory Assistant
Privacy-first RAG pipeline running entirely offline. Phi-3 (4-bit quantized) via llama.cpp on CPU-only hardware with hybrid dense-vector and keyword retrieval.
Edge Vision Pipeline
Real-time voice pipeline with NVIDIA NeMo STT and custom wake-word detection. YOLOv8 models trained and deployed for human tracking, package classification, and gesture control at the edge.
HEXAPOD
Multi-legged locomotion system with ROS2 control stack deployed on Raspberry Pi via Docker. Geometric IK for deterministic real-time execution across six coupled limbs.
6-DOF Robotic Arm
Full motion planning pipeline with inverse kinematics, trajectory execution, and collision avoidance. Simulation-validated before hardware deployment.
SENTINEL
Offline mesh emergency communication using ESP-NOW — zero infrastructure dependency. Autonomous fall detection and gas hazard sensing with mesh-propagated alerts.
Retail Performance Engine
Store performance forecasting with XGBoost and lag-based feature engineering. K-Means segmentation across 50+ retail locations for supply chain optimization.
StockMetrics Pipeline
Multivariate regression on 20 years of Big 5 IT firm data. ETL pipeline aligning daily volatility with quarterly financials. EBITDA Margin Change as statistically significant predictor.
Stock Correlation Platform
Time-series alignment with timezone normalisation and market calendar sync. Pearson / Spearman / Kendall correlations with Fisher z-transform confidence intervals.
Chess Platform
Real-time multiplayer chess with Firebase Firestore synchronization. Versioned state updates prevent race conditions. Stockfish engine offloaded via FastAPI for deep position analysis.
Stock Correlation Platform
Time-series alignment with timezone normalisation and market calendar sync. Pearson / Spearman / Kendall correlations with Fisher z-transform confidence intervals.
Memory Assistant
Privacy-first RAG pipeline with React frontend and FastAPI backend. Automated ingestion from PDFs, DOCX, images via OCR, and web pages with optimistic UI updates.
ROS · ROS2 · MoveIt · Gazebo · RViz · ROS2 Control · URDF
ESP32 · Raspberry Pi · Arduino IDE · ESP-NOW
PyTorch · YOLOv8 · SDXL · LoRA · NVIDIA NeMo
Sentence-Transformers · llama.cpp · ChromaDB
scikit-learn · XGBoost
Python · C++ · TypeScript · SQL
Docker · FastAPI · Linux · Git · SQLAlchemy · Firebase · React
Have a question?
Ask anything about Mitanshu's experience, skills, or availability.
Open to research engineering, ML engineering, and robotics software roles — full-time or internship. Responses within 24 hours.