Software Engineer

Gubin Geramifard

Toronto
Selected Work
SignD landing page
Live in production
Applied AI · RAG · Full-stack

SignD Contract Explainer

A retrieval-augmented AI that reads leases and contracts, auto-flags risky clauses, and answers plain-language questions with the exact clause cited. Cohere embeddings, reranking and generation run over a per-user Postgres and pgvector store behind Supabase auth. A benchmark showed reranking lifted top-1 retrieval accuracy by 17.5%. Python, FastAPI, Next.js, Supabase; deployed on Render and Vercel.

SignD automatically flagging risky lease clauses SignD answering a question with cited source clauses SignD feature overview SignD account sign-up screen
Tally three-way ledger benchmark
Live in production
Systems · Go · Concurrency

Tally Ledger Engine

A double-entry ledger engine in Go that can't lose or duplicate money under concurrency. A single-writer state machine with idempotency and a write-ahead log; a live benchmark shows a naive version conjuring millions of dollars while Tally stays exact to the cent. Group commit makes durable writes ~13x faster than fsync-under-a-lock, and a sharded engine scales the write path across cores with cross-shard transfers made atomic via a two-phase clearing protocol. Verified under the race detector with property-based and fuzz tests, plus CI and a single-binary Docker deploy. Go standard library only.

Tally dashboard: the ledger always sums to zero Tally dashboard overview
ember optimization benchmark: 10.8 to 98.6 tokens per second, a 9.1x speedup
Systems · C++ · Performance

ember Inference Engine

A CPU inference engine for transformer language models, built from scratch in modern C++20 with no ML libraries. It runs the full Llama‑2 forward pass and generates text on the CPU. I optimized it 9.1x (10.8 to 98.6 tokens/sec) with a hand‑vectorized AVX2 kernel, a hand‑rolled thread pool, and int8 quantization; profiling showed the workload was memory‑bandwidth‑bound, so quantization drove the biggest gain, at a 0.4% accuracy cost. Verified correct against a reference implementation, with a unit test suite and CI.

Gubin Geramifard
A little about me

Hi, I'm Gubin!

I'm a software engineer based in Toronto who likes building things people actually use. I have loved video games my whole life, from console to PC, and it was sandbox games that first pulled me into coding. Having the freedom to build anything I could imagine was addictive, and that same urge to create is what led me from life sciences at McMaster into computer science.

Most recently I built production AI features during my internship, from a report-generation engine to a speech-to-text pipeline, and I care most about making sure the things I ship actually hold up. Outside of code, you will usually find me gaming, at the gym, hanging out with friends, or playing with my dogs.

Experience
Software Engineer Intern @ ThePsychoEd Clinic
Sep 2025 to Aug 2026
My toolkit
Languages
PythonTypeScriptJavaScriptGoC/C++JavaSQLDart
Frameworks
ReactNext.jsNode.jsFastAPIFlaskFlutterThree.js
Tools & Cloud
DockerCMakeGitGitHub ActionsAzureGoogle CloudVercelRenderPostgreSQL
Contact

Let's work
together. Say hello →