Backend developer in Ahmedabad, India. I spent 2025–2026 writing Python against the Odoo ORM for client ERP deployments — computed fields, onchange logic, approval workflows, scheduled jobs, and the PostgreSQL queries that stop being fast once a client's data gets large.

I come at the same systems from the other side as well. An ERP is a machine for producing records, and records can be wrong on purpose.

erp-anomaly-bench — a labeled event-log benchmark for ERP fraud detection. Rather than simulate a process, I drove a real Odoo instance (Docker, plus a custom audit addon) through procure-to-pay flows and injected seven fraud typologies into them: duplicate invoices, three-way-match violations, split purchases, subtle overbilling. Every trace and every event comes out carrying a ground-truth label.

Eleven detectors benchmarked against it under a leakage-safe temporal protocol, from audit rules and Isolation Forest up to two deep sequence models (a denoising autoencoder and a BINet-style model). Four typologies are easy — one document is visibly wrong. The other three are invisible to any method that looks at a single trace at a time, deep models included: they score at or near random.

The finding is what closes that gap. Widening the feature scope — time-of-day features and a ±14-day requester×vendor context window — takes a plain Isolation Forest from 0.099 to 0.970 AUPRC on threshold-dodging split purchases, and 0.536 to 0.947 overall, finishing roughly 0.5 AUPRC ahead of both deep models. Same estimator; only what it was allowed to see changed. One typology — a 1–3% skim hiding inside legitimate billing variation — resists every method, and ships as the benchmark's open challenge.

signspeak — two-way Indian Sign Language translation. Webcam signs are read as MediaPipe hand landmarks and spoken through a tiered TTS stack, with Vosk handling speech-to-text for the hearing side's reply. The pipeline is landmark-only by design; no video is ever stored. The dynamic word-sign tier runs MediaPipe Holistic into an LSTM over 30-frame sequences.

Built to a production standard rather than a demo one: 8.6 ms p50 inference on CPU, a 62-test headless suite, GitHub Actions CI, and an evaluation protocol reporting per-sign F1 and cross-signer (leave-one-signer-out) accuracy.

Python · Odoo ORM · PostgreSQL · Django · scikit-learn · OpenCV · MediaPipe · Pandas · NumPy · Docker · pytest · GitHub Actions · Linux

B.Tech in Computer Engineering, L.J. University, 2022–2026. Ahmedabad, India. Reachable at aminmiral0007@gmail.com.