Built for real constraints.

Amazon caps product titles at 75 characters, and if yours is longer its own model quietly rewrites it — dropping the details a seller was careful about, while the seller stays legally responsible for the result. Listing Intelligence reads the live page and runs four narrow research agents in parallel (competitors, keywords, reviews, compliance rules), then a synthesiser writes the title and a separate adversarial agent red-teams the draft before anyone sees it. Every agent is handed a schema instead of being asked to reply in JSON, cheap models do the legwork while the strong one writes the output that ships, and character counting happens in plain code — so the tool cannot do the one thing it exists to prevent.

Most driver-monitoring systems answer one question — is the driver drowsy right now. Saarthi answers that with a deterministic reactive lane, then adds a predictive lane that forecasts at +2, +5, and +10 seconds and refuses to speak unless the evidence supports it. Face landmarks become a small, interpretable time series; a causal buffer keeps the model honest about not seeing the future; Shift-Gated Conformal Abstention and explicit sensor-health checks decide whether a forecast is supportable, and every meaningful decision is written to an audit log. It ships behind a virtual CAN seam so the integration path is real without touching a vehicle. A research POC, not a homologated safety component — and it says so.

When a vehicle downloads a software update, the process has to survive dropped connections, low battery, and interrupted power without ever leaving the car in a broken state. I designed a state-machine controller that makes secure OTA updates resumable and cancellable at any point, with the transport encrypted end to end. It runs on production Jaguar Land Rover systems.

A decentralized exchange built to understand how financial infrastructure works from the inside. Smart contracts hold the trust and enforce the rules; automated liquidity pools set prices and handle the flow. Users swap ERC-20 tokens on Ethereum with no intermediary — the contract itself is the counterparty.

An async task-orchestration service built for scale. It sustains 1,000+ concurrent users with real-time notifications and runs 40% faster than the baseline it replaced. Building it was a lesson in how systems actually behave under load — where queues back up, where latency hides, and how to design for graceful degradation.

A serialization layer for real-time data flowing between Vehicle Stability Control and Automatic Braking. This is communication that has to be deterministic and correct every single time — the failure mode is a car that doesn't stop. Built on FreeRTOS with raw sockets and IPC for predictable, low-latency transfer.

ResNet101 and VGG16 implemented from scratch in PyTorch, reaching 92.2% accuracy classifying 275 bird species. Beyond the numbers, it was a study in how pattern recognition works — how stacked convolutions learn edges, then textures, then whole concepts, whether in silicon or in the visual cortex.

An end-to-end encrypted communication layer for headlamp actuators that improved communication efficiency by 10%. A small percentage that compounds into real impact when multiplied across every vehicle on the road — the kind of low-level optimisation that never shows up in a spec sheet but matters at scale.