Recruiter-facing, source-grounded Agentic Digital Twin with bounded planning, tool use, and claim verification

Hold Space and speak: a 70 MB Windows tray app that types your voice into any field. Whisper Large v3 on Groq's free tier, phrase-streamed as you talk.

Reproducible benchmark of security-oracle construct validity on 140 real CVE fixes.

Production RAG with continuous retrieval evaluation and drift gating — fails CI when retrieval quality regresses against a committed baseline.

Evidence-first portfolio: trustworthy AI for software engineering, software security, and reliable agents.

Research profile: trustworthy AI for software engineering, software security, and reliable agents.

Customer-side GitHub Action for signed, hash-pinned PatchPilot API migration campaigns.

Reproducible study of secure-coding instruction placement for AI coding agents

Five typed, bounded MCP servers for grounded CV, repo intelligence, web research, ATS jobs, and durable outcomes

Effect Browser runs one-query research and browser tasks through durable, action-bound approval and crash-recovery controls that avoid blind retries after ambiguous external effects.

Production RAG and tool-calling backend: adaptive retrieval routing, hybrid BM25 + vector search fused with RRF, MMR diversity, and grounding verification that reports whether the answer is actuall…

Event-driven multi-agent coordinator with role-based fan-out, at-least-once delivery, retries, dead-letter recovery, causal traces, and partial-result fan-in.

Durable, resumable execution for tool-calling AI agents: event-sourced journal, deterministic replay, human approval gates, and crash recovery that never re-fires side effects.

Reachable conservatively analyzes Python call graphs to classify vulnerable code as REACHABLE, NOT_REACHABLE, or UNKNOWN and emit OpenVEX or SARIF evidence.

Evidence-backed ledger for post-sale commitments with human approval, explicit supersession, accepted fulfillment evidence, and a hash-chained audit trail.

Self-hostable LLM inference gateway: policy routing with fallback chains, circuit breakers, semantic caching, per-tenant cost accounting, and Prometheus metrics.

TrustDesk drafts cited vendor-security questionnaire answers from local policies, requires human review, and blocks export of unapproved or stale answers.

Effect Broker executes agent side effects through stable intent, explicit idempotency contracts, authoritative reconciliation, and unknown outcomes instead of unsafe retries.

Agent Redteam runs authorized adversarial probes and reusable guardrails for LLM agents, producing evidence-backed verdicts, risk scores, and regression baselines.

Local-first personal-memory prototype with encrypted storage, retrieval, and source-linked answers.