I work on two sides of the same problem: AI systems (medical imaging, multimodal fusion, LLM tooling) and the robots that run them (ROS 2, RL locomotion, sim-to-real). Most of what's below either ships as open source or is headed to a paper.
🔒 A few research repos below are private until their papers are submitted or published — links will go live with the papers.
trajlens — ruff for robot data
Lints LeRobotDataset datasets (v2.0–v3.0) against their own declared metadata — 16 checks across structural, semantic, temporal, statistical, and video categories, with terminal / JSON / HTML / SARIF reports that drop straight into CI. Validated by auditing a random sample of 100 public Hugging Face Hub datasets: two known upstream lerobot bugs were found in ~19% and ~3% of successfully linted datasets.
bodhirax — local-first LLM prompt eval studio
A visual studio for testing and comparing prompts across Ollama, OpenAI, Anthropic, Groq, and Hugging Face — runs entirely on your machine, keys never leave localhost. Exact / fuzzy / embedding / LLM-judge metrics. Final checks in progress; launching soon.
| Project | Contribution |
|---|---|
| deepset-ai/haystack + core-integrations | 4 merged PRs in the Haystack LLM orchestration framework |
| JdeRobot/RoboticsAcademy | 16 merged PRs — fixed an FP16 crash in the object-detection pipeline, refactored the Hardware Abstraction Layer, shipped 52 unit tests |
| Work | Area | Status |
|---|---|---|
| MRI Reconstruction — dual-branch physics-guided deep learning; learned gating routes k-space adaptively, no anatomy labels at inference. +1.78% SSIM, 112 ms/slice | Medical imaging | Conference paper under review — IEEE SPL · Dr. Anurag Tiwari, TIET |
| ThermoBridge 🔒 — bidirectional 3D MRI↔CT synthesis via Brownian-bridge diffusion with a learnable 3D anisotropic-diffusion mixer | Medical imaging | Paper in preparation · Dr. Anurag Tiwari, TIET |
| Multi-Modal Sensor Fusion for SLAM in Visually Degraded SAR Environments — pipeline-oriented survey of 55 systems (2011–2025) | SAR robotics | Survey submitted · Dr. Ankit Soni, TIET |
| SAR rover — simulated multi-modal (audio-visual-thermal) search-and-rescue platform 🔒 | SAR robotics | Journal paper #2, in progress · Dr. Ankit Soni, TIET |
| Parkinson's early detection — CNN on voice + DNN on MPU9250 tremor, late fusion 88% → 91%, ESP32 sensor-to-inference pipeline | Multimodal ML | ELC summer research, TIET |
Canary Rover — autonomous mine-inspection scout · capstone
PPO-trained locomotion (200K timesteps, PyBullet) → ROS 2 sensor stack (IMU / RPLiDAR / 4× BLDC encoders) → full Isaac Sim 5.1 simulation with live sensor feeds. Currently in sim-to-real transfer; physical rover roughly half built. Team of 5.
object_finder — map-less warehouse AMR · ▶ demo video
ROS 2 Humble skid-steer robot that navigates without SLAM: FSM + PD wall-following on 360° LiDAR, HSV + depth target tracking, tested in the AWS Small Warehouse Gazebo world.
| Project | Description |
|---|---|
| Archon | • Natural-language brief → live deployed site: hybrid RAG (Cohere dense + BM25, RRF fusion) feeds grammar-constrained code generation • Celery async workers, Redis Pub/Sub live WebSocket logs, GitHub App deployer with short-lived tokens, full Prometheus/Grafana/OTel observability |
| Axon | • Fully local tri-modal assistant (text/speech/vision): BART-MNLI zero-shot router dispatches to knowledge RAG, confirmed OS tool execution, or chat • Hybrid retrieval (Qdrant dense + BM25) reranked by a cross-encoder; GBNF grammar-constrained tool calls; zero external LLM APIs |
| TruthTag | • Classical CV pipeline cross-verifying RFID FASTag claims against physical vehicle geometry: 3780-dim HOG features + LinearSVC, 97% accuracy on 24K+ images • Cross-modal centroid tracker with MOG2 virtual tripwire, misclassification error analysis, and a Streamlit operator audit dashboard |
Languages
Robotics & Simulation
AI & Computer Vision
Backend, Data & Infra
