Nishant D.

Top Rated Plus ML engineer: medical imaging (segmentation, registration, nnU-Net), clinical NLP, deep computer vision. 100% Job Success · 1,200+ hrs. I take on the problems where the obvious approach fails. WHERE I DO MY BEST WORK Medical imaging ML. I evaluate and debug segmentation models (Dice scores, Bland-Altman analysis), run rigid and deformable image registration (ANTs/SyN), handle nnU-Net preprocessing and training, and deliver 3D Slicer overlays your clinical team can actually inspect. On my longest imaging engagement I sent a demo video at every milestone, so the client never had to guess where things stood. Before Upwork, I worked on chest X-ray segmentation and calcification detection for a US medical-device software company. Clinical & scientific NLP. 600+ hours (and counting) on a research-heavy clinical NLP project: sentence embeddings, UMAP + HDBSCAN clustering, custom composite scoring, synthetic training data. The kind of work where you form a hypothesis, run the experiment, and let the data kill it. (Under NDA) Deep computer vision. I trained a Flux DensePose ControlNet from scratch — dataset prep, captioning, GPU training, parameter tuning — then shipped it behind a Streamlit app so the client's non-technical team could generate images without me. Also: YOLO detection, and an automated color-correction pipeline that detects a color card and generates 3D LUT/CUBE files across EXR, CR2, CR3, and PNG workflows. Document intelligence & LLM pipelines. I build document-processing pipelines that extract structured data from messy PDFs, scans, invoices, and reports — OCR, transformer-based page classification, and LLM extraction with Claude/OpenAI — delivering clean CSV or JSON. I've also built cross-referencing systems that connect information across multiple documents, and raised extraction accuracy on a pipeline other teams had given up tuning. I ALSO GET CALLED WHEN SOMETHING IS SILENTLY BROKEN Two of my favorite projects were rescues: a Temporal Fusion Transformer backtest that was silently leaking data (fixed the encoder/decoder windowing, ran leakage checks, rebuilt walk-forward validation), and a multi-timeframe trading pipeline writing empty tags (root-caused it, patched it, added audit scripts and trace columns). If you need someone to debug and fix an ML model, training pipeline, or backtest that's producing numbers nobody quite trusts, that is a job I know how to do. HOW I WORK Fast replies — usually within hours, working Nepal time with real overlap across US, EU, and AU hours. Milestone demos, often as short videos. Documentation your next engineer can pick up cold. All 13 of my completed contracts are five stars, and several reviews use the phrase "above and beyond", that part I'm proud of. Send me the problem you suspect is too messy for a freelancer. Those are the ones I want.