Available · Applied ML / AI Engineer

Benyamin
Gidanian

Applied ML/AI engineer in Los Angeles building end-to-end systems, from raw data to deployed, tested inference services. Now expanding into NLP, LLMs, RAG and MLOps, shipping a real project at each step.

bgidanian@gmail.com Resume ↓
0.9557
Macro F1 · 27 PII types
0.9887
AUROC · lens finding
522k
Recipes · hybrid search
71
Tracked runs · 11 backbones
↓Scroll
01

Try the model

— the PII redaction API, live
Token classification · DeBERTa-v3

Hi, I’m Benyamin Gidanian. I fine-tune models in Los Angeles and ship them to production.

Sample sentence above. Your text goes to the live API.
27 entity types · public demo is limited to 10 requests a day Full demo ↗
02

Selected work

— shipped, measured, tested
Live in production

PII Redaction API

Three DeBERTa-v3 models (base, small, xsmall) fine-tuned on 27 PII entity types, served as a FastAPI service on AWS ECS Fargate with a custom domain, Cloudflare TLS, and a GitHub Actions pipeline that lints, type-checks, tests, builds and deploys every push.

0.9557
Macro F1 · base
DeBERTa-v3FastAPIDockerECS FargateCloudflareGitHub Actions
In progress · RAG

Recipe Genome

Hybrid retrieval (BM25 + vector) over 522,517 recipes in Weaviate, embedded with BAAI/bge-small-en-v1.5 through a GPU-backed sidecar so index-time and query-time vectors come from the same model.

522,517
Recipes indexed
WeaviateBM25bge-small-en-v1.5Docker
ML4SCI DeepLense · GSoC 2026 evaluation

Gravitational Lens Finding

Binary lens detection under 16.6:1 class imbalance with ConvNeXt-Tiny, within 0.006 AUROC of a purpose-built equivariant network (Parul et al., NeurIPS ML4PS 2024), using a general backbone and careful imbalance handling instead.

0.9887
AUROC · binary
ConvNeXt-TinyEfficientNet-B0timmPyTorch
Inference service · 102 classes

Flower Classification API

A PyTorch Lightning pipeline across 11 backbones, 71 runs tracked in MLflow, and the project where I found a data-leakage bug that had silently inflated every run’s validation and test accuracy to ~1.0.

0.9837
Accuracy · ViT-B/16
LightningMLflowFastAPIHF HubPyright
03

Agent-built apps

— directed with Claude Code

I set the specs, architecture, invariants and test requirements, and I reviewed and verified each change. Claude wrote the code.

Hybete

Hybrid running and lifting app for iOS and Android: GPS runs, gym logging and multi-week plans, local-first with per-user Supabase sync under row-level security. An AI coach proposes changes into an inbox the user accepts or dismisses; an OAuth 2.1 MCP server lets users connect their own assistant.

FLUTTER · FASTAPI · SUPABASE · MCP · 170+ TESTS · PRIVATE

Invoicer

Multi-tenant B2B invoicing and inventory SaaS for small wholesalers, covering the full buy-and-sell loop. FIFO costing on an append-only stock ledger with row locking; requirements recovered from a 1990s desktop ERP and 186k real documents.

DJANGO · POSTGRESQL · HTMX · 327 TESTS · PRIVATE

Mise

Recipe app that stores LLM-generated recipes as JSON validated against a JSON Schema. The import prompt embeds the live schema, so model output can’t drift from what the validator accepts.

FLUTTER · JSON SCHEMA ↗

Baker’s Calculator

Bread calculator rewritten around a pure-Dart domain layer: baker’s percentages, inverse flour solving for sourdough, scaling and backward-scheduled bake timelines. Every output of the original version is a golden test.

FLUTTER · DART · RIVERPOD ↗
04

Stack

ML / AI
PyTorch · PyTorch Lightning · TensorFlow · Keras · CUDA · Hugging Face Transformers & Hub · Torchvision · timm · ConvNeXt · BERT · Sentence Transformers · Scikit-learn · Reinforcement Learning
NLP / LLM
LangChain · RAG pipelines · Vector databases (ChromaDB / Pinecone) · RAGAS evaluation · Prompt engineering
MLOps
W&B / MLflow · Evidently AI · Model registry · Experiment tracking · FastAPI · Docker · AWS (ECR · ECS Fargate · SSM) · Cloudflare · GitHub Actions · Pytest · Ruff · Pyright · uv
Data
NumPy · Pandas · Feature engineering · Data preprocessing
Web & mobile
Django · Flask · React · htmx · Alpine.js · REST APIs · PostgreSQL / SQLite · Flutter / Dart · Riverpod · Supabase
Agentic development
Claude Code · Claude API · MCP
Languages & tools
Python · SQL · JavaScript · C · Git · Linux / Bash · Jupyter
05

How I work

01

Understand first

I break systems apart mentally before touching code. Knowing why something works matters as much as that it works.

02

Build to learn

Small, practical experiments teach me more than tutorials.

03

Ship it tested

CI pipelines, typed code and reproducible setups are part of the work, not afterthoughts.

04

Iterate on results

I benchmark, measure, and go again. 63% → 93% doesn’t happen in one shot.

Foundation
Machine Learning Specialization
Stanford University & DeepLearning.AI
Supervised & unsupervised learning, neural networks, decision trees, recommender systems
Harvard CS50
Harvard University
C, Python, SQL, algorithms, AI search & logic, web with Django/React
Now
Building the Recipe Genome retrieval eval harness, and looking for ML and applied AI engineering roles.
06

Off the clock

— hybrid training, bread, guitar and big questions
Training week
SunFull body
MonHard run
TueUpper body
WedHard run
ThuLegs
FriEasy run
SatRest
Runs: intervals, long runs, easy Lifting: a Liftosaur program, usually 2 sets per exercise
59:59
10K target · anything under an hour
1
First pull-up

Baking

Sourdough, challah, croissants, focaccia and gelato.

Guitar

Beginner, playing in the Carlebach style.

Big questions

Philosophy (P vs NP, AGI, the problem of evil), neuroscience and self-development.

Markets

I follow AI infrastructure. My working thesis is a sector rotation:

GPUsMemoryInference chipsPhotonics & power
07 — CONTACT

Let's build something that ships.