Faisal Malik Widya Prasetya

ExpertCrowd SWE Benchmark, built on the HUD platform. Designed and shipped RL environments and verifiable tasks used to train and evaluate frontier coding agents — owning each task end to end, from reproducible sandbox setup through grader design, validation, and deployment. Tasks were contributed to Toptal's ExpertCrowd SWE Benchmark dataset, built and evaluated on HUD's RL environment platform.

• Built containerized, deterministic task environments (Docker) pinned to real-world repository snapshots, ensuring hermetic, reproducible rollouts for agentic RL training and evals.
• Wrote task specifications calibrated for clarity and target difficulty, iterating on prompt–grader misalignment to close off reward hacking and specification gaming.
• Engineered programmatic graders and verifiers (hidden unit tests, weighted rubric checks, blocking criteria) producing reliable reward signals with minimal false positives and false negatives.
• Authored reference solutions (gold patches / gold.diff) plus negative controls (no-op and lazy/partial solutions) to prove graders reward genuine task completion only.
• Evaluated each task against multiple frontier LLM agents over repeated rollouts, analyzing pass@k, trajectories, and failure modes to land difficulty in the target pass-rate band.
• Enforced quality bars before deployment: solvability, grader robustness, contamination/leakage checks, and reviewer QA.

Impact: Runner-up contributor to Toptal's ExpertCrowd SWE Benchmark dataset, ranked on accepted, high-quality RL tasks.

Global commercial biobank & human biospecimen CRO for biopharmaceutical and academic researchers.

• Took an end-to-end ML system for biospecimen analysis (semantic segmentation, cellular segmentation, heatmap classification) from concept to production on AWS — owning data labeling and dataset curation, training, evaluation, serving, and inference.
• Fine-tuned H-Optimus-0, an open-source pathology foundation model, using parameter-efficient LoRA/PEFT adapters (Hugging Face Hub and timm) to specialize frozen ViT embeddings for downstream segmentation and classification heads without full model retraining.
• Designed scalable, fault-tolerant serverless inference on AWS (Lambda, SQS, DynamoDB, S3, ECS, Fargate) for highly parallel, tile-level model serving on whole-slide pathology images.
• Built the high-performance backend for the customer-facing product pathboxai.com in TypeScript (Bun runtime, ElysiaJS).
• Built MLOps CI/CD automation for continuous experimentation, training, and hyperparameter tuning via GitHub Actions workflow dispatches that instantiate and orchestrate on-demand EC2 runners.
• Implemented image transformation techniques that improved visualization quality for researchers and lifted downstream model performance.
• Diagnosed bottlenecks in the core image processing pipeline and rewrote it in Rust, significantly cutting inference-path latency and operating cost.

• Rolled out AI coding agents (Claude Code) on development servers with scoped permissions and authorization, aligning AI tooling with the client's access-control and data-usage policies.
• Built a conversational agent that interprets user intent, plans multi-step actions, and executes them autonomously (Pydantic AI), driving browser and desktop applications through Playwright and Pywinauto.
• Led design through launch of a production autonomous trading platform — 50+ specialized agents, 28 subsystems, and 20 microservices — trading live on a £140,000 FTMO prop account.
• Integrated MetaTrader 5 as the gateway layer for trading signals, order execution, and account state on live accounts.
• Built a fail-closed governance layer that constrains agent behavior at runtime — 10 canonical rules, 12 sequential validation gates, HMAC-SHA256 envelope signing, and strict separation of duties.
• Architected an immutable, hash-chain-verified audit trail across 500+ sealed files, with tamper detection, automated audits every 6 hours, and a bypass detector.
• Hardened security with AES-256-GCM encryption at rest (Windows DPAPI key binding), loopback-only network binding, file integrity monitoring, and automated secret scanning in CI/CD.