class Sailesh:
def __init__(self):
self.role = "Autonomous Systems & Deep Learning Engineer"
self.school = "Amrita Vishwa Vidyapeetham, Amritapuri"
self.focus = ["ADAS", "Sensor Fusion", "Vision-Language Models"]
self.stack = ["Python", "PyTorch", "CARLA", "Docker"]
self.building = "Dual-Perception planning for autonomous driving"
self.learning = ["Diffusion Policies", "3D Geometric DL", "CUDA"]
def philosophy(self):
return "Ship the pipeline, not the demo."- 🚗 I build closed-loop autonomous driving stacks in CARLA — perception, fusion, safety, planning, control
- 🧠 Currently researching VLM-guided trajectory vocabularies with Hydra-MDP style multi-head scoring
- 🌪️ Also into geometric deep learning — neural mesh deformation for aerodynamic drag reduction
- ⚡ LeetCode Knight · 1894 rating · Top 4.79% globally
- 💬 Ask me about YOLOv8, LiDAR–camera projection, PointNet, LoRA fine-tuning
|
Vision-language planning stack that pairs a trajectory vocabulary (64 prototypes, 4s horizon) with Qwen3 + CLIP perception. Candidates are ranked by a four-head Hydra-MDP scorer — imitation, collision, drivable area, comfort.
|
Neural mesh deformation that lowers a car's drag coefficient without re-running CFD. A PointNet surrogate predicts Cd; a deformation net nudges vertices under gradient descent. Benchmarked against GraphSAGE, MeshCNN, DGCNN.
|
|
Full ADAS + semi-autonomous stack at a sustained 20 FPS. 32-channel LiDAR→camera projection, YOLOv8s + YOLOP panoptic perception, speed-scaled panic bubble, Frenet planning, occlusion-robust AEB.
|
Decision-intelligence ADAS built on YOLOv8 + UFLDv2. Drops Kalman/EMA smoothing for a zero-latency perspective mapping algorithm; TTC engine from RGB-depth fusion drives FCW and AEB through a deterministic arbitration matrix.
|
|
LoRA-tuned FLAN-T5 classifies text as SENSITIVE/SAFE, then a NetworkX pipeline scores node-level risk with PageRank and strike counts to flag repeat spreaders in child-facing communities.
|
Two offline, Dockerised PDF engines. TextSense extracts document outlines through a multi-signal heuristic (structural > stylistic) with no ML at all. InfoWeave ranks paragraph chunks against a persona query using MiniLM embeddings and cosine similarity.
|
If a pipeline only works in the demo video, it doesn't work.