Gray Scrubs AI, continued

• Built an RGB-only, role-labelled 3D digital twin of the operating room from five ceiling cameras: fine-tuned SAM 3 (848M) for OR panoptic segmentation on one GPU, patching its upstream training code, and fused it with feed-forward multi-view reconstruction. Benchmarked it against 7 published 3D methods on depth-sensor ground truth; an ablation showed the domain fine-tune carries the semantics.
• Fine-tuned LTX-Video 13B (LoRA) on 645 multi-camera OR clips for mask- and depth-controlled video generation. Now simulating physics-labelled rare-event training video in Isaac Sim, and reproduced the OmniFall fall-detection baseline on AWS GPUs and the UK's Isambard-AI supercomputer (GH200, Slurm), for a CVPR 2027 paper in preparation.
• Designed an LLM labelling pipeline (LLM labellers, blind judges from two model families, human review in self-hosted CVAT) that produced 6.7K hand-contact labels; fine-tuned the HOI-DETR hand-object detector on them and extended it into an RGB-only sterile-field touch detector, configured on validation before one held-out test read.
• Earlier experiments: streaming surgical-phase recognition (frozen DINOv3 and V-JEPA 2 features, causal HMM decoding), PaliGemma 2 vision-language QLoRA for procedure coaching, and a multitask model exported to ONNX for Jetson.
• Built the clinical data side on AWS (Terraform): a hospital video-ingestion platform (recorder-disk reader, resumable S3 upload, GPU screening workers, a passkey-protected FastAPI/React review portal) and DICOM de-identification (OCR quarantine of burned-in text, NER scrubbing of reports, CT defacing) with Lambda/S3 portals for surgeons and data partners.

Arantic Digital, continued

• Primary engineer of a generative-AI marketing content app in production since Jul 2026: LLM generations run as BullMQ jobs with live SSE progress, timeouts and stalled-job recovery; thread history with refine and fork in 20 languages; file attachments passed to the model as context.
• Wrote the provider-agnostic LLM layer (Vercel AI SDK; Anthropic, OpenAI, Mistral) with a live model catalog, cost estimates and encrypted keys, plus a prompt editor whose layouts compile to Zod schemas for structured output.
• Designed the real-time notification system: Server-Sent Events fed by a BullMQ worker in its own container, with cross-tab leader election so each user holds one connection; later merged with AI progress into one multiplexed event stream.
• Implemented a recursive-descent search query parser and SQL builder over PostgreSQL full-text search (weighted tsvector, GIN), and a product module with trigram search and a reusable transactional CSV import.

Contact Center Specialists, continued

• Ran production infrastructure (FreeSWITCH telephony, PostgreSQL, MongoDB, Redis, Zabbix and Grafana monitoring, call-record billing); parallelised the Go device-provisioning service and built a real-time voice-AI flow on Gemini Live.

SCIT CAIR, continued

• Remote-sensing research with Dr. Usman Nazir: ViTs and CLIP filtering on 600K+ satellite images (NeurIPS 2024 paper).

Beaconhouse National University, continued

• Built a Next.js/tRPC LMS with role-based access and SSO, and a geospatial analytics dashboard for HBL Bank.