Index of work
Systems that had to keep running after launch: inference serving at twenty million requests a day, forecasting under real money, model training pipelines, clinical risk scoring. Some shipped to production, some stayed research, all of them ran against real data.
2–3 s
Response time · 10M+ downloads
Vyro · 2023
Triton Inference Server deployment for a photo enhancement app at consumer scale, with batching tuned so a single GPU could hold the interactive latency budget.
80%
Less time to ship a feature
Vyro · 2023
One Python interface that hosts any Stable Diffusion workflow, so shipping a new generative feature became a config change rather than a deployment project.
−40%
Processing latency
Add Life · 2025
MediaPipe wired into Unity and tuned for mid-range phones, holding steady under ten concurrent video streams after stress testing and memory-leak work.
2 modes
Inline diffs or queued repos
Personal · 2025
FAISS retrieval supplies repository context, an LLM writes the review, GitHub Checks delivers it. Runs inline for small diffs or through a Celery and Redis queue for larger repositories.
0.024
RMSE, best GRU model
Research · 2024
A deep learning pipeline for multi-horizon financial time-series forecasting, comparing RNN, GRU and LSTM families across horizons rather than tuning a single favoured architecture.
4
Architectures, grid-searched
Research · 2023
A modular PyTorch framework for benchmarking CNN architectures, ResNet-18, MobileNetV2, GoogLeNet and AlexNet, with grid search across optimizers and hyperparameters so the comparison is like for like.
ROC-AUC
Scored with SMOTE-corrected classes
Research · 2023
An end-to-end binary classification pipeline for clinical risk detection, with SMOTE correction for class imbalance and confusion-matrix reporting alongside the headline score.
85%
Urdu voice command accuracy
Personal · 2021
Home automation driven by spoken Urdu commands, recognised by a CNN trained on a self-collected dataset, with a React Native front end for the appliances themselves.
20M
Requests a day · 99.5% uptime
Vyro · 2022–2024
A bespoke inference layer behind a top-two US AI art generator. Thirty-plus models scheduled across Kubernetes, with the request queue and the GPU pool tuned together rather than separately.
Live
Catalogue-grounded retrieval
Stealth startup · 2024–2026
Type a dish and get a recipe whose ingredients are semantically matched to real supermarket products, then added straight to the cart. RAG over a live catalogue, not a static ingredient table.
+45%
Search relevance
Stealth startup · 2024–2026
Keyword lookup replaced with transformer embeddings in Elasticsearch, fronted by a normalization pipeline that cleaned messy retailer product names before anything was indexed.
15 min
Turnaround, down from hours
Vyro · 2023
Docker-based avatar fine-tuning on Runpod and AWS, treating GPU capacity as something you rent per job rather than per day.
4 stages
Read, judge, tailor, draft
Personal · 2026
A session-based agentic copilot for job applications. LangGraph runs the reasoning, PostgreSQL with pgvector holds semantic memory across chat threads, so it reads a posting, judges fit, tailors the resume and drafts the outreach without losing the thread.
End of index
The five worth reading are written up in full.