Fiscora  ·  ANKA  ·  PlateHunter

I’m Sam Afshar. I started programming around 2024, beginning with Python. Since then, the interesting part has shifted from “can I train a model?” to a harder question:

Can I turn the idea into a complete system that handles data, uncertainty, validation, failure states, interfaces, and real constraints?

That is where I spend most of my time now: artificial intelligence, machine learning, computer vision, software engineering, and product-oriented technical experiments.

I’m still early in the journey. I would rather make that visible than cover it with inflated titles, fake skill percentages, or performance claims that the underlying work cannot support.

Fiscora is a competition-oriented personal-finance intelligence application designed around a simple separation: use explicit algorithms where inspectability matters, and use learned models only where prediction is genuinely the experiment.

The repository currently includes:

• a deterministic month-end projection built from recent spending and weekday behavior, including bootstrap-based uncertainty;
• an explicit financial-health score, anomaly checks, recurring-payment detection, savings scenarios, and a What-If simulator;
• receipt OCR with review before transactions are committed;
• a multi-task LSTM that consumes a six-month sequence of scale-independent financial features to estimate next-month spending growth and overspend probability;
• an experimental Conv1D/LSTM Bitcoin model evaluated against simple baselines rather than presented as a trading oracle;
• a FastAPI backend, local JSON persistence, bilingual Iran/US behavior, notifications, and a context-aware AI assistant that retrieves computed financial facts instead of inventing account context.

The model-evaluation story is intentionally not polished into marketing. The behavior model’s recorded results come from a synthetic-but-coherent training panel, not a production population. The Bitcoin notebook’s documented future-dated synthetic test also does not establish a material forecasting advantage over naive baselines. I think keeping that visible makes the project stronger, not weaker.

Engineering focus: Python · FastAPI · TensorFlow / Keras · LSTM · Conv1D · EasyOCR · pandas · NumPy · HTML / CSS / JavaScript

Source → github.com/SamAfshar2011/innoverse2026-Fiscora

ANKA is an earthquake-safety competition project built as a preliminary decision-support and educational system, not a replacement for a structural engineer.

A user enters eight building and seismic parameters. The backend derives additional engineered features, aligns them with the training schema, scales the numerical inputs, and feeds a TensorFlow dense network that returns a collapse-risk estimate. The surrounding application then turns that estimate into understandable context rather than stopping at one number.

The current repository contains:

• four risk bands from low through critical;
• rule-based safety recommendations in English, Persian, and Turkish;
• a browser-based 2D earthquake simulator with its own time-stepped physics loop and visual damage state;
• a contextual AI assistant that can receive the building state;
• generated PDF reports plus a report-registry / verification workflow;
• explicit disclaimers that predictions, recommendations, simulation, and reports are informational and preliminary.

The important boundary is part of the engineering: ANKA is not a certified structural inspection, engineering report, or governmental safety evaluation.

Engineering focus: Python · TensorFlow / Keras · feature engineering · HTML / CSS / JavaScript · multilingual UI · PDF/report workflow · LLM-assisted guidance

Source → github.com/SamAfshar2011/1i1w_competition

PlateHunter is an Iranian automatic license-plate-recognition pipeline built as several explicit stages instead of one opaque “OCR model.”

The inference path is:

frame → YOLO plate detector → plate crop → YOLO character detector → 8 sorted character crops → ConvNeXt classifier → format/confidence validation

For valid Iranian plates, the implementation expects DD L DDD DD: seven digit positions plus one letter/symbol at zero-based position 2. A result is not silently accepted just because each model returned something. The pipeline checks the exact character count, plate structure, stage-level confidence thresholds, and a final geometric-mean confidence so one weak stage is harder to hide.

Video adds another engineering layer: an 8D constant-velocity Kalman filter, Hungarian assignment, appearance/geometry/OCR matching signals, and track-level OCR consensus to keep plate identities stable across frames. The inference configuration also includes an Apple MPS path for supported macOS hardware.

The repository publishes the training/inference code and evaluation artifacts, but it explicitly does not include the private training dataset or trained model weights.

_{Copied from the project’s own public inference assets; not a recreated mockup.}

Engineering focus: Python · PyTorch · Ultralytics YOLO · ConvNeXt Small · OpenCV · SciPy · Kalman filtering · Hungarian matching · Apple MPS

Source → github.com/SamAfshar2011/Persian_ALPR_project

I don’t use percentage bars for skills; they create precision without evidence. These are the technologies and problem areas that show up in the work above:

Python · TensorFlow / Keras · PyTorch · YOLO · ConvNeXt · FastAPI · OpenCV · scikit-learn · SciPy · pandas · NumPy · OCR · HTML · CSS · JavaScript · REST APIs · multilingual interfaces · Apple MPS

I’m especially interested in the boundary between AI systems and products: where model behavior meets software architecture, human-readable feedback, and decisions that have to remain understandable.

2024 — Started learning programming with Python.

Then — Moved from individual scripts and models toward machine learning, computer vision, APIs, frontends, and complete application pipelines.

Now — Building larger systems, documenting their real limitations, and trying to raise both the engineering and product-quality bar with every project.

The long-term goal is larger than a portfolio: build a technology startup around products that are genuinely useful, then learn how to turn it into a serious company. I’m not there yet. That is the direction.

I closely follow the work of Apple, NVIDIA, Microsoft, Google, and Anthropic, and I’m interested in the company-building and technology careers of Steve Jobs, Bill Gates, and Sam Altman.

This is not a logo wall or a claim of similarity. They are simply part of the technology landscape I study while forming my own taste in products, engineering, and ambitious companies.

Most of the interesting detail is in the repositories themselves: notebooks, validation logic, architecture decisions, limitations, and the parts that did not produce a perfect result.

Explore my GitHub → github.com/SamAfshar2011

_{Sam Afshar · born December 2011 · learning by building}