Ryan Junejo
Projects
I'm interested in making AI systems easier to evaluate, understand, and control. Here are some things I've built toward that. More on GitHub.
AI Safety
Trojan Harbor
An indirect prompt-injection evaluation suite for Harbor: eight matched attack/clean task families, deterministic graders, and a documented admission process.
The frozen suite ran 560 attempted trials across seven models (559 valid). Attack success ranged from 0% for GPT-6 Astra and GPT-6.1 Sol to 35–40% for the GPT-5.6 models and 50% for GPT-6 Luna.
EvidenceGraph
A research prototype for reconstructing what AI agents actually did during an incident.
EvidenceGraph reconciles agent transcripts against independently collected records, separates supported actions from ambiguous or contradicted claims, and produces an auditable evidence docket with citations back to the original source data.
In the Apart Research AI Incident Response Sprint report, a receipt-relay failure showed that possession of a genuine receipt is not enough for event attribution. After the fix, twelve paired evidence conditions passed 36/36 contract checks. On a staged registry incident across ten seeds, it correctly supported 9/9 production attributions, contradicted both fabricated receipts, and made no confident errors.
Agent Mafia
A deterministic multi-agent evaluation environment for studying strategic deception and social reasoning in language models.
I ran 12 model endpoints across 40 games and 4,267 agent turns, using replayable event logs and game-state ground truth to evaluate claims independently of whether an agent ultimately won. In the primary analysis set, roughly 118–120 of 433–460 distinct claims were verifiably false.
Products
Crashlabs
A policy-driven testing platform for catching AI agent failures before they reach users.
We convert system prompts into Policy Graphs, generate adversarial multi-turn scenarios, and simulate realistic users through a Simulation Orchestrator. Teams run Stress Tests and handwritten Test Suites against voice agents and get crash-test reports in minutes.
Crashlabs · Technical whitepaper
Sunday
A production voice and email operations platform I built at Crash Labs, for fitness and wellness businesses.
We deployed Sunday across 27 locations and reached $100k in contracted ARR. The agents handled roughly 700 calls and more than 1,400 emails per week, with 60% inbound containment. They answer the phone, draft inbox replies for staff review, and run booking workflows inside the customer's system.
Research & Evaluations
When Should a Coding Agent Use a Bigger Model?
An experiment measuring the marginal value of escalating a coding agent to a more capable model partway through a task.
I forked live SWE-bench trajectories at matched checkpoints, holding the filesystem, conversation, and remaining budget fixed, and compared continuing with the base model against escalating. Across the sampled checkpoints, escalation increased solve probability by roughly 0.4.
Biodiversity Classification
Multimodal ML research I worked on at the Vector Institute.
I built and evaluated systems for species classification, incorporating vision-based approaches and comparing performance against models including OpenAI CLIP and Meta Hiera.
Systems
PokeSpeed
A deterministic environment for building and evaluating AI agents that speedrun Pokémon Red and Blue.
It includes RAM-grounded observations, replay verification, milestone tracking, pathfinding, and an LLM reference agent. The longer-term goal is to make game completion time a measurable optimization target for increasingly capable agents.