Skip to content
View professor3333's full-sized avatar

Highlights

  • Pro

Block or report professor3333

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
professor3333/README.md
Niroj Danai, Machine Learning Engineer. End-to-end ML systems, measured on data they have never seen.

I build machine learning systems from data pipeline to deployed API, then test them on data they have never seen. When a model misses its target, the README says so.

Bachelor of Engineering "Software Engineering", Nepal College of Information Technology, 2026  ·  Kathmandu, Nepal

Open to Machine Learning Engineer and Applied ML Engineer roles, remote or on-site.

Featured work

ProtocolLedger: 0.766 entity F1 on WNUT-2020, against 0.703 for a CRF baseline. Relation F1 0.580 against 0.504. ProtocolLedger  ·  Live demo
Turns free-text lab protocols into an editable prep sheet, with every value linked to its source text.
Fraud Risk Scoring: 0.907 ROC-AUC on a held-out later month, at 43% lower cost than approving everything. Fraud Risk Scoring  ·  Live dashboard
Calibrated fraud scores that approve, review or block within an analyst team's daily budget.
WildInbox: 0.447 macro-F1 on nine unseen cameras, against 0.285 for the frozen baseline. Grouping alone cuts the review queue by 61.4%. WildInbox
A review inbox for trail-camera photos: groups shots into capture events and suggests a species for each.
PassageWatch: 0.213 count error (nMAE) across four unseen locations, against 0.388 for a tuned classical baseline. Lower is better. PassageWatch
Counts fish moving upstream and downstream in sonar video, keeping the evidence behind every count.
Shelf Life: 75% of its daily shortlist were real removals on held-out days, 6.8 times the unaided hit rate. Shelf Life  ·  Live app
Predicts which job postings come down within seven days, trained on a panel I scraped from seven job boards.
BeltWatch, in development: 0.336 foreground macro IoU with SegFormer-B0, against 0.152 for a tuned random forest. BeltWatch
Segments cardboard, plastic and metal contamination on paper-recycling conveyors and queues uncertain images for review.

How I work

Five steps: data, train a simple baseline then the real model, test on unseen data, ship an API with a live demo, and send uncertain cases to a person. Corrections feed the next training run.

Toolbox

Python PyTorch scikit-learn XGBoost Transformers OpenCV ONNX Runtime MLflow DVC FastAPI Streamlit PostgreSQL Redis Docker GitHub Actions AWS

Pinned Loading

  1. ProtocolLedger ProtocolLedger Public

    🧪 Evidence-linked extraction of actions, materials, quantities & conditions from lab protocols: span + relation NLP with human review and traceable JSON/CSV export.

    Python

  2. wildinbox wildinbox Public

    Review app for trail-camera photos: groups a memory card into capture events, suggests the species with a fine-tuned EfficientNet-B0, and sends uncertain events to a person. FastAPI, Streamlit, Pos…

    Python

  3. passagewatch passagewatch Public

    Directional fish-passage counting for sonar video. Detects and tracks fish, counts upstream and downstream passages, and lets a reviewer correct the counts and export a traceable report. Evaluated …

    Python

  4. beltwatch beltwatch Public

    Deep-learning contamination auditing for paper-recycling conveyors: semantic segmentation, visible-coverage estimates, uncertainty-aware review queue, and auditable human feedback.

    Python

  5. fraud-risk-scoring fraud-risk-scoring Public

    Fraud risk scoring on IEEE-CIS: temporal validation, XGBoost, calibrated thresholds, FastAPI serving

    Python

  6. shelf-life shelf-life Public

    Predicting how long a job posting stays on the board: an end-to-end classical ML system built on my own scraped job-listing panel.

    Python