▸ /identity
node 0x01
02 //
research signal
◈
decision context — behavioral memory for llm agents
forge_dc compiles human corrections and self-assessed tool failures into imperative RULE directives and injects them at the decision points where they apply. dc-bench — a benchmark for behavioral memory: 71 real-trace scenarios across 9 dimensions. accepted at the neurips 2026 palm workshop.
95%
compliance
+85
point gain vs. no memory
▲ from 10% → 95%
10.7×
fewer injected tokens
159 vs. 1,705 naive retrieval
71
benchmark scenarios
9 behavioral dims
+46
points vs. best prior system
memp 49% → forge_dc 95%
03 //
field operations
active
▸Senior AI Engineer · Agentic Systems
PwC · Toronto
Building a self-maintaining, self-improving, self-evaluating agentic framework — agents that assess their own output, learn from their failures, and carry that learning across sessions. Taking behavioral memory from research into production systems.
completed
▸AI Research Engineer · R&D
Georgian · Toronto
Building behavioral memory for llm agents (forge_dc, dc-bench), the platform's MCP server (8 production integrations), OpenGeorge context engine, and a multi-agent blog-post pipeline. Discovered the "only rules rule" finding — published at neurips 2026 (palm).
completed
▸Graduate Researcher · GNNs
York University · Elder Lab
Leading a graph-based framework for contour grouping — fusing CNN features with multi-step message passing for local + global shape context. Built a synthetic dataset of 100k+ shapes for pretraining. Distributed PyTorch + Hugging Face Accelerate.
archived
▸Teaching Assistant · Python
York University
Taught and mentored 200+ students across multiple programming courses; led labs and one-on-one sessions. Introduced test-driven development and algorithm tracing into the lab curriculum.
archived
▸AI Research Intern · Spiking Neural Networks
University of Tehran
Built the DeepSNN framework in PyTorch for simulating spiking neural networks — CNN layers, pooling, neural-encoding methods. Earlier life: bachelor's research under the bio-inspired computation umbrella.
04 //
projects
P-01
Nutrition Fact AR Reader
Real-time AR system; YOLOv8-seg fine-tuned to 95% mAP@50. PaddleOCR + local LLM (Ollama) for structured parsing. Unity AR overlay with comparison & recommendations.
yolov8 · paddleocr · unity · ollama · flask
P-02
BART PEFT Multitask
Fine-tuned BART with PEFT for joint summarization + sentiment classification under a shared encoder-decoder. Deployed via Flask API and Gradio for real-time inference.
bart · peft · hugging face · flask · gradio
P-03
FoodSeg103 · Mask2Former
Fine-tuned Mask2Former on FoodSeg103 (103 categories) via transfer learning + hyperparameter tuning, boosting mIoU. Interactive Gradio app.
vision transformers · semantic seg · hugging face
P-04
Cross-Modal Image Retrieval
Cross-modal image-text retrieval combining VGG16 visual features with Word2Vec embeddings, aligned via cosine similarity over 10,000+ images.
pytorch · vgg16 · word2vec · cv+nlp
P-05
FEAT · Few-Shot Learning
Implementation of FEAT (Few-Shot Embedding Adaptation) with a custom meta-learning framework and data loaders; validated on standard few-shot benchmarks.
pytorch · meta-learning · tensorboardx
05 //
tooling readout
— languages
python
c / c++
sql
bash
hcl
assembly
— ml / ai
pytorch
tensorflow
hugging face
pydantic-ai
langgraph
langchain
opencv
scikit-learn
— agentic ai
multi-agent
tool use
mcp
claude · gpt · llama
lora / peft
rag
behavioral memory
prompt compilation
benchmark design
— mlops / infra
docker
kubernetes
terraform
ci/cd
faiss
pinecone
chromadb
mlflow · w&b
opentelemetry
langfuse
— cloud
gcp
vertex ai
cloud run
bigquery
aws
bedrock
azure ai
— dev tools
claude code
anthropic sdk
mcp
sqlalchemy
alembic
slack bolt
apscheduler
uv
06 //
interactive terminal
nv-shell — bash 5.2 · session 0x7f2a
↑history tabcomplete /focus
nima@nv:~$