Hexapod
18-DoF hexapod with a full ROS 2 stack. Wrote the URDF xacro model and a complete ros2_control hardware interface from scratch. Dockerised the runtime with NVIDIA GPU support.
ROS 2 Humble, ros2_control, Pinocchio, Docker
Mitanshu Goel is a robotics and AI/ML engineer based in Delhi, India. ECE graduate from MAIT (2026) with an AI/ML minor. Currently building a bimanual VR teleoperation rig at Variety Innovation / Enferent.ai. Meta Quest 3 drives an Elite Robots CS66 pair and a Franka Research 3 over a real-time Linux control loop. He has run six continued-pretraining experiments on a self-scraped Reddit corpus (Mistral 7B, Qwen 2.5, nanoGPT) across four hardware tiers, with artefacts kept private. He sole-authored the ros2_control hardware interface and Ignition Fortress migration for an 18-DoF hexapod at atom-robotics-lab. Looking for Robotics SWE, Research Engineering, ML Engineering, or Applied AI roles.
Delhi, India · Actively looking
Mitanshu Goel
Building robots and training models.
Focus
Most of what I do sits at the intersection of hardware and learned models. Right now that means a bimanual VR teleoperation rig — a Meta Quest 3 driving an Elite Robots CS66 pair, with the same control loop being extended to a Franka Research 3 so it works across arm families. The rig double-duties as a dataset collector: every teleoperation session logs synchronised arm states and headset pose for downstream policy training.
On the model side, I've been running continued pretraining experiments on a Reddit corpus I scraped and processed myself. Six runs across Mistral 7B, Qwen 2.5 at a few scales, and a nanoGPT I built from scratch — partly to ship experiments, partly just to understand what training actually feels like from the ground up.
Experience
Robotics SWE & Physical AI InternMar 2026 – Present
Variety Innovation / Enferent.ai · Gurugram, On-site
- Building bimanual VR teleoperation infrastructure. Meta Quest 3 drives an Elite Robots CS66 pair, extended to a Franka Research 3 so the system covers both arm families.
- Implemented damped-Jacobian IK via Pinocchio — manipulability-adaptive damping, null-space regularisation, and three singularity guards on the critical joint and TCP rotation axes.
- Built the imitation-learning dataset recorder. It captures synchronised left and right arm states, VR headset pose, and TCP poses across both platforms. The company's main goal is to ship multi-robot teleoperation datasets at scale, and we train policies on top of the data we collect.
AI & Robotics InternJun – Aug 2025
SarthakAI · Delhi
- Built a real-time voice pipeline with NVIDIA NeMo's FastConformer-Transducer STT, added custom wake-word detection, and integrated it into a UBTech Yanshee humanoid.
- Trained and deployed custom YOLOv8 models for three production tasks: human tracking, package classification, and gesture-based robot control. Each one had its own training regime and inference pipeline.
- Designed a fault-tolerant interface between the robot hardware and the AI inference agent — polling-based, built to keep running even when hardware responses are intermittent.
- Built an ESP32 / Raspberry Pi telemetry workstation bridging embedded firmware with Python analytics layers.
Robotics InternJul – Sep 2024
Nextup Robotics · Delhi
- Configured a 6-DOF arm in ROS/Gazebo — debugged URDF kinematic configurations and tracked down the sim-to-real discrepancies that were blocking stable trajectory execution.
- Integrated MoveIt for inverse kinematics and collision-aware trajectory planning in C++. Cut execution time by roughly 50% after switching planners and tuning parameters.
Education
B.Tech, Electronics & Communication Engineering2022 – 2026
Maharaja Agrasen Institute of Technology (MAIT) · Delhi
- Minor: Artificial Intelligence & Machine Learning
Projects
Robotics
6-DOF Robotic Arm
Motion planning pipeline for a 6-DOF arm — IK, collision-aware trajectories, and sim-to-real validation. The hard part was tracking down URDF kinematic mismatches that kept blocking stable execution.
ROS, MoveIt, Gazebo, C++, Python
Sentinel
Emergency mesh communication that works with zero infrastructure — just ESP32 nodes talking to each other over ESP-NOW. Fall detection via MPU6050, with gas sensing packed into the same firmware. Mesh propagates alerts automatically.
ESP32, ESP-NOW, C++, MPU6050
AI & ML
Reddit CPT
Six continued pretraining runs on a Reddit corpus I scraped and processed myself. Mistral 7B at two ranks, Qwen 2.5 at three scales, and a nanoGPT built from scratch to understand what training actually feels like end to end. Full pipeline from raw Pushshift data to packed tokenization.
Unsloth, TRL, PEFT, LoRA, rsLoRA, Mistral 7B, Qwen 2.5, nanoGPT, A100
Darwin Studio
What if image generation worked like evolution? Darwin Studio treats SDXL latent tensors as genetic material — mutation, crossover, generations. A custom moment-preserving SLERP handles the interpolation and keeps child latents on the correct noise manifold so they decode properly.
PyTorch, SDXL Lightning, Diffusers, custom SLERP
Memory Assistant
An offline personal memory store with hybrid dense and keyword retrieval. Phi-3 runs locally via llama.cpp for metadata at ingest time. The /api/v1/ask endpoint is the real RAG path — retrieve memories, generate a cited answer, return it with source references. No cloud required.
FastAPI, ChromaDB, Sentence-Transformers, Phi-3 GGUF, llama.cpp
RAG Assistant
A RAG system where hallucinations are structurally impossible by design. If the model claims "Fully Answered" but has no citations to back it up, Pydantic rejects the response at parse time — before it ever reaches the user. Built with a retry loop that feeds parse failures back to the model, and an eval that found 0 hallucinations across 9 test questions.
Llama 3.3 70B, Groq, ChromaDB, Pydantic
Data Science
RetainIQ Churn Model
Churn prediction using a Cox proportional hazards model, with six risk signals extracted from Steam review text by an LLM. Careful about leakage — explicitly removed covariates that encode the event label. Hold-out C-index: 0.874 vs 0.640 without the LLM signals.
Cox PH (lifelines), Groq Llama 4 Scout, scikit-learn
StockMetrics Pipeline
Tested 25 fundamental-variable pairs across 5 Indian IT firms (2005–2025) for predictive power on annual stock returns. One pair came out significant — Wipro EBITDA margin change, p=0.029. The other 24 didn't, and the README says so plainly. Null results are results too.
pandas, scikit-learn, F-test, statsmodels
Stock-Influence Platform
Upload any time-series and see how it correlates with a stock price. Three correlation methods (Pearson, Spearman, Kendall), each with proper 95% Fisher z-transform confidence intervals. The math is in the code and can be checked.
FastAPI, React, pandas, SciPy, yfinance, Chart.js
Web
Chesstra
Real-time multiplayer chess with Firestore. Server decides who wins — if the moving player crashes mid-move, the opponent's client detects it and writes the resolution. Presence heartbeats with a 15s liveness window. Stockfish runs as a separate FastAPI service with a cold-start wake-up ping.
React 19, Firebase Firestore, FastAPI, Stockfish, Vite
Technical Stack
Robotics
ROS 2 Humble, ROS, MoveIt, Pinocchio, ros2_control, RViz.
AI & ML
PyTorch, Unsloth, TRL, PEFT, LoRA, rsLoRA, BitsAndBytesConfig, Hugging Face Hub.
Mistral 7B, Qwen 2.5, nanoGPT, SDXL Lightning, Diffusers, NVIDIA NeMo, YOLOv8.
Sentence-Transformers, ChromaDB, llama.cpp.
Data Science
lifelines (Cox PH), scikit-learn, XGBoost, K-Means, pandas, NumPy, SciPy, statsmodels.
Fisher z-transform CIs, F-tests, NDCG, MRR, Recall@k.
Backend & Web
FastAPI, Uvicorn, SQLite FTS5, BM25.
Next.js 15, React 19, Firebase Firestore, TypeScript.
Docker, Linux, Git.
Languages
Python, C++, TypeScript, JavaScript, SQL, Bash.
ASK
Have a question?
Ask anything about Mitanshu's experience, skills.
Contact
Looking for full-time roles in Robotics SWE or Physical AI. Also open to Research Engineering and applied ML.
Phone+91 85956 57583
GitHubgithub.com/mitanshu-2004
LinkedInlinkedin.com/in/mitanshugoel
LocationDelhi, India. Open to relocation.