class Soham:
def __init__(self):
self.role = "Robotics + AI Engineer (in training)"
self.education = "Dual Degree (B.Tech Mech + M.Tech Robotics) @ IIT Madras"
self.cgpa = 8.75
self.focus = ["Deep Learning", "Reinforcement Learning",
"Motion Planning", "Real2Sim", "NLP"]
self.philosophy = "Decide → Build → Stay consistent → Evolve"
def current_obsession(self):
return "Making robots and policies that behave correctly in the real world"- 🤖 Dual-degree student at IIT Madras building at the intersection of robotics, ML, and systems
- 🛠️ Comfortable going from-scratch — Transformers, neural nets, RANSAC planar fitting, RL pipelines
- 🏭 Summer Intern @ Procter & Gamble (supply-chain & packing-line optimisation)
- 🎯 Targeting roles in AI/ML, robotics, and quant/analytics
- 📫 Reach me: me22b191@smail.iitm.ac.in
⚠️ Make these repos public and replace#with the real links so they show up here.
| 🔬 Project | What it does | Stack |
|---|---|---|
| Transformer for Neural Machine Translation | Transformer (Vaswani et al.) from scratch — multi-head attention, sinusoidal PE, Noam schedule, BLEU eval on Multi30k (DE→EN) | PyTorch |
| Multi-Task Vision Pipeline | Unified VGG11 backbone with 3 heads — classification, bbox regression (custom IoU loss), U-Net segmentation on Oxford-IIIT Pet | PyTorch, W&B |
| Information Retrieval System | TF-IDF VSM baseline + BM25, LSA (SVD), WordNet expansion on the Cranfield benchmark; metrics from scratch | Python, NLTK |
| Multi-Robot UR5 Motion Planning | Goal-biased RRT for up to 5 UR5 arms (30-DOF composite C-space) with per-step collision checking | PyBullet |
| LLM-Agent Meeting-Scheduling Env | OpenEnv-compliant RL benchmark — FastAPI + Pydantic + Docker, LLM-agent baseline with reward shaping | FastAPI, Docker, HF |
| Classical ML — From Scratch | PCA / Kernel PCA, K-means, spectral clustering, EM, regression, + a full NN with 6 optimisers — no ML libs | NumPy |
| Autonomous Mobile-Robot Stack | ROS 2 nav node: A* global + RRT local + pure-pursuit, LiDAR avoidance, ArUco TF, pick-and-place (e-Yantra Stage 2) | ROS 2 |
| Swarm Drone Search-and-Rescue | 10-drone FSM swarm, sector exploration, gap-scan recovery under No-GPS / No-Comm / Kill-zone constraints | Webots |
