Case Study
AI Audit Platform
AI-driven smart contract auditing platform: isolated Docker sandboxes, a multi-phase Claude-powered hunt pipeline, Foundry PoC verification, and an orchestrator that scales audit workers.
Problem
Manual smart contract audits do not scale, and AI alone lacks grounding in concrete findings. The system needed isolated sandboxes, a phased hunt pipeline, Foundry PoC verification and adversarial false-positive filtering so findings arrived with evidence — not noise.
My role
Internship contribution at Softstack. Primary focus: Docker sandbox pool orchestration, sandbox ↔ backend bridging, the TypeScript V2 audit phase pipeline and skill selection, EVMBench-style detection measurement, and audit UX surfaces (timeline, memory, resume). Softstack owns the production system; the public repo is an architecture and contribution write-up with original illustrative samples.
Architecture
Technical decisions
Isolated Docker sandboxes
Each audit runs in a pooled, isolated container with Foundry, Slither, Mythril and Claude tooling. That keeps toolchains reproducible, contains side effects and lets the orchestrator scale workers independently of the API.
TypeScript phase pipeline
Migrated audit execution toward a typed, phase-aware pipeline with skill selection and context budgets — clearer than a bash-centric entrypoint when coordinating protocol understanding, static analysis, AI hunt and PoC writing.
Foundry PoCs against mainnet forks
Candidate findings are verified by writing and running Foundry proofs of concept. Adversarial filtering then tiers evidence so reports prefer grounded vulnerabilities over speculative AI output.
Challenges
Sandbox ↔ product bridge
Job assignment, env injection, progress callbacks and cancellation had to stay reliable across orchestrator, containers and frontend without leaking fragile control paths.
Measurable detection, not anecdotes
Built and hardened an EVMBench-style evaluation path so detection rate and confidence could be measured across pipeline changes instead of judged by feel.
Outcome
A production audit platform developed for Softstack, combining Dockerized sandbox orchestration, a multi-phase AI hunt pipeline and Foundry-backed verification. Portfolio materials document architecture and my concrete contributions without publishing Softstack source.