Rana Muhammad Usman

ACL is a Go-based domain-specific language designed to bring structure, determinism, and auditability to AI agent workflows. The project introduces a contract-first approach where every task is defined with intent, execution steps, required evidence, and a final receipt that validates outcomes. I designed and implemented the full system, including the DSL syntax, execution engine, and validation layer. The key contribution is shifting agent workflows from probabilistic outputs to verifiable processes, enabling reproducibility and debugging in production environments. This makes it comparable to tools like Terraform or Airflow, but purpose-built for AI agents and large language model (LLM)-driven systems.

AgentBridge converts any OpenAPI specification into agent-ready tools within seconds, enabling large language models (LLMs) to interact with APIs in a structured and scalable way. I built the system to simplify tool integration for AI agents by automatically parsing API schemas and generating callable functions. The project reduces integration effort and allows developers to rapidly expose services to AI systems without manual mapping. It demonstrates a practical approach to bridging traditional APIs with modern agent-based workflows.

Bragfeed is a social platform designed around authentic self-expression through achievements, collections, and lifestyle choices, without the pressure of traditional follower-based systems. Instead of likes and vanity metrics, discovery is driven through curated circles and content relevance. I conceptualized and built the platform end-to-end, including product positioning, core UX principles, and back-end logic. The key innovation is redefining “bragging” as a signal rather than noise, enabling users to share high-value moments while connecting with audiences who actually care. The platform explores alternative social graph dynamics and content discovery mechanisms beyond traditional social media models.

HireSignal is an AI-driven system that extracts meaningful hiring signals from unstructured candidate data, such as resumes, job descriptions, and behavioral patterns. The platform focuses on identifying underlying intent, skills, and fit rather than relying on keyword matching. I designed the system to combine natural language processing (NLP) techniques, structured extraction, and scoring models to create actionable candidate insights for recruiters. The core contribution is transforming messy hiring data into structured, decision-ready signals, improving screening efficiency and match quality. The platform reflects my broader focus on turning unstructured data into reliable, usable outputs.

RedBee is an autonomous AI agent that runs through Discord. I built it as a practical “AI employee” that can handle job search, Twitter/X posting, repo promotion, analytics, and scheduled tasks. The system uses an LLM for planning, then runs deterministic Python code for repeated execution, so it does not waste tokens on every cron tick. It includes a Discord control layer, browser-based job application flow, Twitter automation, repo promotion, SQLite-backed memory, and config-driven personalization. The project shows my ability to design AI agents that move beyond chat and actually operate tools, schedules, browser flows, and memory.

Agent WorkGraph turns Claude Code and Codex sessions into reusable engineering memory. I built it to solve a real problem I kept facing: every AI coding session starts from zero. WorkGraph watches local agent sessions, compiles useful lessons into a .workgraph/ directory, and produces artifacts such as journey notes, repo rules, known traps, generated evals, task templates, and reusable skills. It also includes a local web UI that shows milestones, confidence levels, files touched, and a knowledge graph. The project is local-first and designed to help engineering teams reuse what agents have already learned, rather than repeating the same mistakes.

AutoThink is a one-click AutoML library for tabular data. I built it so a user can pass a dataframe and target column, then get a working model with minimal setup. The system detects the task type, validates data quality, handles preprocessing, engineers features, trains LightGBM, XGBoost, and CatBoost models, optimizes ensemble weights, calibrates probabilities, and produces post-training diagnostics. I also added benchmark artifacts comparing AutoThink with FLAML and AutoGluon across multiple datasets and time budgets. This project is directly relevant to predictive modeling work because it reflects my practical approach: build strong baselines, measure clearly, and ship usable models.

TinyForge is a local self-improvement loop for small coding models. I built it to test whether a tiny model could improve by learning from its own failures, without cloud APIs, a teacher model, or human feedback. The model attempts coding tasks, runs tests, studies the exact failure, generates repairs, extracts weak-to-strong repair pairs, and trains a lightweight LoRA adapter. The project runs on a MacBook Air and focuses on measurable test feedback rather than vague scoring. In one reported slice, single-pass public test performance improved from 16/50 to 28/50 after training on 13 self-generated repair pairs.