Systems thatthink at scale.

An enterprise platform that automates candidate screening end to end. Every incoming CV is parsed, embedded, semantically matched against open job descriptions, scored, and ranked — without a recruiter having to open a single profile first.

Scoring every ticket in one pass

Evaluating a candidate against one requisition at a time does not scale with the number of open roles. The scoring engine fans a single parsed CV across every relevant hiring ticket concurrently, which is what turned screening time into a 91% reduction rather than a marginal one.

Keeping LLM calls off the request path

Model latency is measured in seconds and request budgets in milliseconds. Matching and ranking run as asynchronous work with results persisted on completion, so the API stays responsive regardless of how long inference takes.

Structure and flexibility in the same schema

Parsed CVs vary wildly in shape. Modelling them as optimised JSONB in Cloud SQL kept relational guarantees for the fields that matter while leaving the long tail queryable, improving query performance and flexibility by 30%.

A learning platform that ingests PDFs, GitHub repositories, documentation and raw code, then generates interactive material from them: Q&A, guided explanation, and coding exercises that are evaluated automatically by executing the learner's submission.

Executing untrusted code safely

Learner submissions are arbitrary code. Each execution is isolated in a Firecracker microVM on EC2 with its own kernel and lifetime, which reduced execution-level security risk by 80% relative to running submissions in a shared container runtime.

One API over three storage models

PostgreSQL holds relational state, ChromaDB holds embeddings, and Redis absorbs hot reads. 35+ REST endpoints present all three as a single coherent interface so the client never has to know which store answered.

Ingesting genuinely heterogeneous sources

A GitHub repository, a PDF and a documentation site have nothing structurally in common. A normalising ingestion layer reduced them to a shared chunked representation, which is what made 90% content coverage achievable.

A computer vision system for real-time object detection and intelligent surveillance, running CenterNet ResNet-101 on a Jetson Nano across multiple live RTSP streams, with a chatbot layer that answers questions about what the cameras actually saw.

A latency budget set by the hardware

CenterNet ResNet-101 had to hold 34ms inference at 32.4 mAP on a Jetson Nano. The accuracy-per-millisecond tradeoff was the whole design, and it was decided by what the device could sustain rather than what the model could score.

Multiple live streams, one device

Five or more concurrent RTSP streams share a single accelerator. Frame scheduling and tracking state per stream kept multi-object tracking above 85% accuracy without any one camera starving the others.

Making the logs answerable

Ten thousand detection records are a data problem, not an insight. A TF-IDF and cosine-similarity chatbot over the logs let a human ask what happened instead of writing a query.