I'm a Computer Science student from Columbia University who builds backend systems, AI/ML applications, and full-stack tools. I trained as a Machine Learning Fellow through the Break Through Tech AI program at Cornell Tech.

I'm most at home designing event-driven backends, shipping full-stack apps against real databases, and wiring LLMs into practical tools. Recent work sits across software engineering, AI/ML, agentic systems, and generative models — details below.

🎯 Focus: Backend & distributed systems · Machine learning engineering · AI-powered tools · Data analytics

Keeping one character looking like itself across any scene — and measuring whether it actually does

Fine-tunes Stable Diffusion XL with DreamBooth-style LoRA so a single character stays recognizable whether it's in a coffee shop, in space, or in armor. The interesting half is the evaluation: prompt fidelity alone says nothing about identity, so the harness scores both.

• Stack: Python · PyTorch · Stable Diffusion XL · Diffusers · LoRA/DreamBooth · CLIP · DINO
• Approach: Low-rank adapters (a few MB) over a frozen UNet, trained against a rare identifier token with optional prior preservation to stop class collapse
• Highlights: Four-metric eval CLI — CLIP-T for prompt fidelity, CLIP-I and DINO for identity fidelity, plus gen-gen self-consistency across scenes — with measured failure modes documented (pose overfitting, identity drift at high guidance, prompt terms the LoRA swallows)

Parallel LangGraph agents that review GitHub pull requests the way a senior engineer would

An event-driven system that reviews GitHub pull requests with parallel specialist agents for security, code quality, testing, and documentation. It's a hands-on study of the reliability engineering around agents — webhook verification, latency isolation, idempotency, verification gates, and cost control — not just an LLM in a loop.

• Stack: Python · FastAPI · Redis · LangGraph · PostgreSQL + pgvector · Docker · LangSmith
• Architecture: HMAC-verified webhook ingestion → async Redis pipeline → four-agent LangGraph with deterministic aggregation → idempotent posting behind a confidence gate
• Highlights: Unified Postgres data layer with pgvector search, bounded retries with failure classification and attempt limits, and full observability — LangSmith tracing, verified cost accounting, and a daily spend ceiling

A ReAct-loop coding agent that reasons, calls tools, and runs code inside a sandbox

Takes natural-language instructions and acts on your filesystem through a tool system — reading and writing files, searching a codebase, and executing code in a Docker sandbox. Built around the ReAct (Reason–Act–Observe) loop and Anthropic's Claude API, with the load-bearing pieces hand-built for understanding.

• Stack: Python · Claude API · Docker · pytest
• Architecture: ReAct agent loop over a tool registry (write_file, read_file, search_code) with context persistence and history summarization at a token threshold
• Highlights: Docker-based execution sandbox with a code validator and pinned timeout bounds, backed by a pytest suite

Three interconnected Next.js apps on a shared Supabase backend

A production-deployed system for authoring AI prompt chains, generating and rating image captions, and administering the platform. Built over a semester for Columbia's COMSW-4995.

• Stack: Next.js (App Router, TypeScript) · Supabase · Tailwind CSS v4 · Vercel
• The apps: Prompt Chain Tool (authoring) · Caption Rating App (public voting) · Admin Panel (back-office + analytics)
• Highlights: Google OAuth with role-based access gates, a four-step caption-generation pipeline, and a vote-analytics dashboard — all built against a shared, RLS-aware multi-tenant database

I'm looking for full-time roles in:

• Software Engineering — backend & distributed systems
• AI/Machine Learning Engineering
• Data Science & Analytics

Currently going deeper on distributed-systems design, LLM infrastructure, and production ML — and always happy to talk shop with fellow builders.

When I'm not building, you'll find me on the soccer field, at the gym chasing incremental goals, or watching films and series for the storytelling. Staying active and engaged outside of tech keeps me a sharper problem-solver and collaborator.

• LinkedIn: linkedin.com/in/colin-j-emmanuel