What I actually know how to do

Custom modules and training loops, AMP autocast, gradient checkpointing, scaled-dot-product Flash Attention, and profiling for VRAM rather than for FLOPs.

Seen in Image-to-LaTeX Handwriting OCR, ViT Compression Trade-offs, Saturn Labs

Hybrid lexical and dense retrieval, reciprocal rank fusion, contextual chunk embedding, per-tenant collection isolation. Two production-shaped services built end to end.

Seen in Multi-Tenant Legal RAG Assistant, Kurt, an Offline-First AI Tutor

Standard ViTs, EfficientViT and DPT-Hybrid. Attention internals to the point of surgery: token merging by key-cosine similarity, proportional rescaling, entropy-ranked head pruning.

Seen in ViT Compression Trade-offs, Image-to-LaTeX Handwriting OCR

INT8 AbsMax and NF4 quantisation through bitsandbytes, token merging, L2 threshold masking and structural head pruning, mapped against a measured Pareto frontier.

Seen in ViT Compression Trade-offs

Encoder-decoder design, 2D positional encoding, autoregressive decoding, ResNet-18 backbones, and mixed-precision training inside an 8 GB budget.

Seen in Image-to-LaTeX Handwriting OCR, ViT Compression Trade-offs, Saturn Labs

Document loaders, splitters and embedding pipelines; sentence-transformers for retrieval and the Datasets library for corpus assembly.

Seen in Kurt, an Offline-First AI Tutor, Image-to-LaTeX Handwriting OCR

Llama-3 and Mistral served locally through Ollama, with hot-swap across five providers including Gemini Vision, driven by a provider descriptor rather than by branching.

Seen in Kurt, an Offline-First AI Tutor

Hybrid conv-transformer backbones, 2D layout preservation for mathematical notation, and real-time hand landmark tracking with temporal filtering.

Seen in ViT Compression Trade-offs, Image-to-LaTeX Handwriting OCR, Hand-Gesture Drawing Tool, Saturn Labs

Whisper-family ASR through Faster-Whisper with voice activity detection, CUDA-accelerated well past real time.

Seen in Kurt, an Offline-First AI Tutor

Q-learning, policy gradients and deep Q-networks. The emergence AI work was about the simulator RL agents exploit rather than about the agents themselves.

Seen in emergence AI

Scientific computing, async services, and the whole numerical stack. The language every project here is written in.

Seen in Multi-Tenant Legal RAG Assistant, Kurt, an Offline-First AI Tutor, F1 Telemetry Dashboard, Image-to-LaTeX Handwriting OCR

Async API design, dependency-injected session scoping, and fail-closed tenancy enforcement at the router boundary.

Seen in Multi-Tenant Legal RAG Assistant

Day-to-day version control and branch hygiene on a Linux workstation, plus static deployment through GitHub Pages.

210 unit tests running in roughly six seconds with no external dependencies, plus a live smoke suite asserting twenty properties against a real Chroma server.

Seen in Multi-Tenant Legal RAG Assistant

ChromaDB collection design, per-course and per-session partitioning, and metadata filtering with centroid steering.

Seen in Multi-Tenant Legal RAG Assistant, Kurt, an Offline-First AI Tutor

Background ingestion pipelines, solo-pool workers where VRAM forbids concurrency, and subprocess isolation for irreconcilable dependency pins.

Seen in Multi-Tenant Legal RAG Assistant

WebSocket and SignalR protocol work down to the payload: subscribing to a live timing feed and decoding base64 plus zlib frames, and tunnelling a self-hosted API to a static front end.

Seen in F1 Telemetry Dashboard, Multi-Tenant Legal RAG Assistant

Typesetting as a build target rather than a document format: Tectonic compilation, matplotlib mathtext as an offline renderer, and a 1,007-token LaTeX vocabulary as a model output space.

Seen in Kurt, an Offline-First AI Tutor, Image-to-LaTeX Handwriting OCR

Performance-critical routines and algorithm implementation. Comfortable, not fluent.

React SPAs and typed component work. The legal assistant's front end, and this site.

Seen in Multi-Tenant Legal RAG Assistant

Schema design and query work against SQLite in WAL mode for document and job state.

Seen in Multi-Tenant Legal RAG Assistant

3D modelling, assemblies and technical drawings for the Mars rover chassis, suspension geometry and mechanical mounts.

Seen in Vicharaka, IISc

Physics-based simulation environments and synthetic data generation for perception training, built during the Saturn Labs internship.

Seen in Saturn Labs

Environment and task construction on top of Isaac Sim, generating labelled data at a volume real capture could not reach.

Seen in Saturn Labs

Closing the gap between a model trained on synthetic frames and one that has to work on a real camera, which is where most synthetic data pipelines quietly fail.

Seen in Saturn Labs

Robot simulation and sensor modelling alongside the Isaac stack.

Seen in Saturn Labs

Parametric modelling and design iteration alongside SolidWorks on rover subsystems.

Seen in Vicharaka, IISc

Physical integration: taking a subsystem from a machinable drawing to a mounted, wired part that survives contact with terrain.

Seen in Vicharaka, IISc

Asset preparation and scene construction feeding the synthetic data pipeline.

Seen in Saturn Labs

Node graphs, topics and system integration for a rover control stack.

Seen in Vicharaka, IISc

Vector spaces, eigendecomposition, SVD and the matrix calculus every one of these models is written in.

Seen in ViT Compression Trade-offs

Multivariable calculus, gradient-based optimisation, and the analysis underneath backpropagation.

Statistical modelling, hypothesis testing and Bayesian methods, plus the quantile construction NF4 quantisation is built on.

Seen in ViT Compression Trade-offs

Convex optimisation, descent methods and constrained problems, applied to both training schedules and Pareto-frontier analysis.

Seen in ViT Compression Trade-offs

Dependent types and tactic-mode proof, used to carry correctness arguments alongside a collision detection implementation rather than testing after the fact.

Seen in emergence AI

Proving a property holds for every input rather than sampling inputs and hoping. Exact rational arithmetic in place of IEEE-754, so determinism is a theorem and not an observation.

Seen in emergence AI

Interpolation, resampling and error analysis. The uniform five-metre distance grid in the telemetry work is a numerical methods answer to a plotting question.

Seen in F1 Telemetry Dashboard

Formal languages, finite automata and decidability. Coursework, and the vocabulary the Lean work borrows its state machine from.

Shannon entropy as a working tool: ranking attention heads by how unfocused their distributions are, then pruning the flat ones.

Seen in ViT Compression Trade-offs

Abstract algebraic structures and their applications in cryptography and physics. Coursework in progress.