I build the system, then dig under it: transformer internals, code-switched language research, local AI agents, computer vision, embedded hardware, and a physical reasoning robot that sees, asks, refuses, and acts.
/ selected systems
Each card is a system with a claim: what it tests, what it proves, and where it still breaks. Completed work gets proof. Future work stays clearly marked.
Extends the original mBERT vs MuRIL comparison across datasets, scripts, platforms, and label policies to test when Indian-language pretraining actually generalizes.
View research repo ↗A ten-project curriculum from transformer internals to a 6-DOF Raspberry Pi arm with vision, local reasoning, uncertainty questions, and code-level safety gates.
A character-level GPT-style transformer built by hand: embeddings, positional encoding, causal attention, residual blocks, training, perplexity, and sampling.
View repo ↗A local-first macOS agent with planning, tool use, voice input, confirmation gates, and a native state-aware presence layer.
View repo ↗A local referee-call explainer grounded in the IFAB Laws of the Game, using page-aware retrieval and a local language model.
View repo ↗Browser-visible MediaPipe hand tracking with pinch-to-grab window movement, smoothing, deadzones, and velocity-sensitive control.
View repo ↗A PlatformIO diagnostic for raw XPT2046 touch coordinates and pressure on an ESP32 touchscreen, with display and serial output.
View repo ↗Build a decoder-only transformer by hand: embeddings, attention, residual blocks, training loops, evaluation, and generation.
Analyze AG News, establish a TF-IDF baseline, evaluate with confusion matrices, F1 and ROC curves, inspect errors with LIME, and write the result.
Serve the Project 2 classifier with FastAPI and Pydantic, test it with pytest, containerize it with Docker, and deploy it through Railway and AWS Lightsail.
Move from CNN fundamentals and filter visualization to CIFAR-10, ResNet-18 transfer learning, YOLO, and real-time webcam detection.
Implement byte-pair encoding, curate a Simple English Wikipedia corpus, train a BPE transformer, and test scaling and domain adaptation.
Learn MDPs and the Bellman equation, compare PPO and SAC, and train a Kuka iiwa arm in PyBullet using reward shaping and curriculum learning.
Connect a virtual camera to the arm through OpenCV detection, pixel-to-world mapping, homography, planning, inverse kinematics, and closed-loop execution.
Assemble offline speech, vision, local reasoning, structured prompts, ambiguity detection, active learning, and independent safety interlocks before hardware.
Build the Raspberry Pi 5 robot, calibrate its camera, solve and validate inverse kinematics, control six servos, and demonstrate safe failure handling.
Implement and compare PCA, t-SNE, UMAP, k-Means, Gaussian mixtures, expectation-maximization, and support vector machines.
GitHub links appear only when a public repository exists. Project 1 is published; Project 2 is currently being built locally.
/ build trajectory
The through-line is simple: understand the model, learn to evaluate it honestly, connect the components into a system, then put intelligence into the physical world.
/ stack map
The current skill stack is intentionally uneven: strong where projects are done, grey-boxed where the next builds will force depth.
Research pipelines, local tools, computer vision systems, automation, and embedded diagnostics.
mBERT, MuRIL, Moondream, Llama, OpenCV, and from-scratch internals.
FastAPI services, local inference, guarded tool use, ESP32 hardware, and the path toward Raspberry Pi control loops.
/ contact
The current mission: build serious proof, document the failed attempts, and make the path from idea to working system visible.