23 PRs merged, 8 issues opened in Hebbian-Robotics/hflow so far. Highlights:

• #457 feat(snapshot): verify delivered snapshots against integrity receipts
• #429 fix(import): CRC-validate reused landing episodes before stamping receipts
• #411 fix(build-ai): cover completeness knobs in the check version contract
• #403 fix(import): report the collector's success label instead of inventing one
• #389 feat(import): per-episode receipts in prepared-manifest.json

Issues I opened:

• #460 doctor labels a healthy video corrupt when a decoder is unavailable
• #426 The resume path stamps a conversion receipt over reused bytes it never integrity-checked
• #405 Eight LeRobot importer metadata refusals have no test, including the pagination loop guards

See all my merged PRs in HFlow

• 22-year-old engineer currently contributing at HFlow, a YC-backed robotics data platform.
• Merged 10 PRs in my first public week, including a LeRobot to MCAP converter and a silent NumPy data-corruption fix that passed unnoticed by the test suite.
• Work reviewed by Kingston Kuan, co-founder of HFlow (YC S26), available as a reference.
• Built a client-facing website from scratch during my summer Full-Stack Development internship, owning major features from requirements through pixel-perfect React and Tailwind implementation.
• Built a Land Purchase Risk Assessment project for Mireye's Co-founder & CTO, Shashwat Kapoor, who called it "excellent work" and praised the architecture and product thinking.
• Researched and published a hybrid AI resume-screening pipeline using Sentence-BERT and Phi-3.5, evaluated across 8,000 resume-JD pairs.
• Drove a 4-person team from data collection and experimentation through final publication.

• Repatch: Autonomous engineering agent that reproduces bugs, authors tests, applies fixes, opens PRs. 7-step state-machine orchestration (UNDERSTAND → EXPLORE → REPRODUCE → PLAN → EXECUTE → VERIFY → SUBMIT). Structured tool calling, Docker/Nixpacks sandboxing, Map of Truth / Semantic Path Grounding.
• vela-cascade-eval: Deterministic evaluation harness for scheduling agents. 500+ seeded scenarios, 713 conflicts generated, 10 machine-checkable invariants, 1,000 determinism checks (byte-identical replay). Naive baseline: 888 double-bookings, 841 priority violations. Cascade-aware agent: zero safety violations, 121 recovery failures (true resource exhaustion).
• groundedness-engine: Zero-LLM structural verification for AI claims. 4 ordered deterministic checks: citation present → section exists → section retrieved → quote matched. Catches "3 years → 5 years" hallucinations without a single model call. Standalone npm package.
• CAREERLENS: 6-stage semantic resume ↔ JD matching pipeline. SBERT bi-encoder (all-MiniLM-L6-v2) + cosine similarity in 384-dim space. 59-node skill ontology (NetworkX DiGraph) with 2-hop transferability (e.g., Docker → Kubernetes). Implicit skill recovery surfaces 25% more skills vs keyword baselines. 9.1% accuracy gain over tools like Jobscan. Dynamic weight allocation per role type. SHAP-style explainability + PDF report generation.
• Nexa · Live: AI trip planner powered by Gemini. Personalized itineraries from preferences, not generic guides. React/Next.js, Node/Express, Tailwind, Framer Motion.

• Building AI agent systems: autonomous engineering agents, orchestration, tool calling, verification, and agentic workflows.
• Engineering evaluation-first: deterministic experiments, invariants, seeded scenarios, replayable results, and honest failure models.
• Building full-stack products: React/Next.js, Node.js, Python, APIs, databases, deployment, and production interfaces.
• Learning in public: how determinism, verification, and reliable failure handling make AI systems trustworthy.
• Open to roles in AI engineering, agentic software engineering, full-stack, backend, and early-stage startup work.

• Deterministic Agent Evaluation: Why Invariants Beat Vibes
• Zero-LLM Verification: Catching Hallucinations Without Models
• Semantic Resume Matching: Beyond Keyword Overlap