Aryan Ahuja

Video-redaction platform for police body-cam footage used by 150+ agencies. Designed the Python redaction pipeline that pairs LLM reasoning with computer-vision segmentation, fanned out across parallel Cloud Run GPU tasks. Built the REST layer that pulls evidence from five body-cam evidence systems across four vendors (Axon, Motorola CCA, Getac EVM4 and EDM3, LensLock), streaming vendor video through SHA-512 hashing into GCS in O(1) memory. Fine-tuned a PyTorch detector on a custom 86K-image dataset (0.63 mAP@50:95), then reimplemented inference in C++ directly on GPU tensors, cutting per-frame latency from 50 ms to 13 ms and unlocking real-time redaction on 30 fps footage. Pitched the platform to three police departments and turned officer and DA requests into shipped features within days.

Python OCR extraction pipeline (Camelot / PyMuPDF + GPT-4 header matching) that turns multi-page lighting specs into validated JSON, cutting B2B quotation time 64% per client at Lumens. Automated four internal workflows for Dell and Cricket Wireless in n8n with API and webhook integrations, cutting manual data entry 70% across 5+ departments.

A hyper-local fishing-intelligence agent that returns cited, condition-aware recommendations at sub-2s p50 latency. A four-node LangGraph pipeline with Langfuse tracing and cost-aware model routing sits over a hybrid Qdrant retriever (1536-d dense + BM25 sparse vectors with recency decay, >0.75 Recall@5) and five per-species XGBoost models tuned with Optuna and tracked in MLflow. A FastAPI SSE backend on TimescaleDB is fed by Celery pollers on nine NOAA stations, with Redis caching and Postgres fallbacks for graceful degradation.

Real-time restaurant-availability alerts. A distributed event-driven platform on an Apache Kafka spine polls 50+ live sources and sustains 400+ events/day at p95 under 60 s end-to-end, with exactly-once delivery via a Lua-atomic Redis scheduler. The operator UI in TypeScript/Next.js streams live status over Server-Sent Events, persisting to a PostgreSQL/TimescaleDB event store behind Docker-containerized async microservices.

Automated highlight detection for long body-cam footage, built in a couple of hours. It embeds every frame with CLIP ViT-L/14, takes cosine deltas between consecutive frames to score visual change, finds the salient moments with a rolling-MAD peak detector, and cuts duration-budgeted clips around them.

Fine-tuned T5/BART to translate U.S. government text into plain language with a self-built 20k-pair corpus, a staged training curriculum, and custom constraint losses (+7.1 SARI).