Neil Mahure
Applied AI Engineer
San Francisco Bay Areaneil.mahure@gmail.com(310) 499-3143linkedin.com/in/neilmahuregithub.com/croresnos
Builds agentic AI systems from LLM primitives rather than framework assembly: provider orchestration, hybrid retrieval, and MCP tool integration. The current build is an operating layer between the user and the host machine, where model output becomes a system action only after it clears a per-tool authorization gate, so the action executed is the action approved. ~261,000 lines across 8 repositories, on a delivery background in enterprise finance systems and go-to-market operations.
AI Systems & Engineering Projects
Agentic Personal AI Operating System (J.A.R.V.I.S.)
Python · Rust · FastAPI · Claude / Gemini / DeepSeek · MLX · LanceDB · SQLite · Next.js · Tauri
- Placed a persistent agent layer between the operator and macOS, executing host actions through one gated surface
- Built the agent router on raw LLM primitives rather than a framework, dispatching by task type across 6 providers
- Routed 31 MCP tools through one authorization chokepoint, so every host write clears a per-tool permission check
- Cut symbol resolution to a graph lookup, pairing LanceDB vector recall with a filtered AST dependency graph
- Tied every phase release to a requirement-id trace over the full suite; the last complete run stood at 2,113 passing
- Traced 2,122 tests to 627 requirement ids, so no phase ships with a requirement nothing exercises
- Wrote ~111,000 lines across 683 commits in Python, Rust and TypeScript, under unit, integration and mutation suites
- Ran adversarial evaluation and the voice interface on-device under MLX on Apple Silicon, with no network call
- Kept the stack self-contained, running DeepSeek-R1, MLX-Whisper, Kokoro TTS and Openwakeword locally
- Specified later phases (sandboxed execution, a Rust authority kernel, multi-device operation) as designed, not built
LLM-Orchestrated Desktop Assistant, Exposed as an MCP Server (F.R.I.D.A.Y.)
Python · FastAPI · OpenAI · Pydantic · OAuth 2.0 · WebAuthn · Faster-Whisper · Twilio
- Shipped a desktop assistant into daily use by 1 real user, across Gmail, Calendar, Tasks and WhatsApp
- Exposed the assistant as an MCP server, so another agent drives it through the same authenticated tool surface
- Implemented Google OAuth 2.0 with scoped consent and refresh-token handling, plus WebAuthn passkey authentication
- Rendered cards, charts, tables and navigation in a Tauri client from Pydantic structured outputs
- Reached daily use in 5 weeks, at 57,723 lines and 579 tests across 129 commits, for one non-technical user
- Frozen at v1.0 deliberately. The authority problem it surfaced outweighed the next feature and became the gateway below
- Added Faster-Whisper speech-to-text and Twilio call and SMS alerting, so voice commands return live context
Local MCP Authorization Gateway with a Durable Effect Ledger (VERONICA)
Python · MCP (stdio) · SQLite · HTTP approval surface · pytest
- Interposed a local gateway on an agent's MCP tool calls, holding every external effect for human approval
- Refused any call rewritten between the approval card and the wire, recording approved intent beside executed fact
- Proved the gated path end to end against a live Claude Code session and a real upstream server
- Surfaced 10 defects in the first live session against a green unit suite, then tested the class the suite structurally missed
- Disclosed the reversal class before each call, failing closed to irreversible wherever nothing can classify one
- Credential custody is designed and not built. The gateway and the ledger run; the vault is the next thing owed
- Extracted from the assistant above at v1.0, on the finding that an agent's report and its effect are different objects
- Carries 3,877 lines and 67 passing tests across 17 commits, as a working slice rather than a shipped product
Environment Preflight Validation Library
Python · pytest · TOML
- Halted execution on 19 manifest checks before first import, turning a late runtime traceback into a startup refusal
- Scoped it to refuse rather than repair, so a failed check never leaves a half-corrected environment behind it
- Reached feature-complete at 150 tests across 7,979 lines, installed from source rather than published to PyPI
Multi-Agent Trading Research Platform (Stocker & Gambler)
Python · PostgreSQL / pgvector · XGBoost · scikit-learn · Alpaca · Kalshi · Polymarket
- Coordinated specialist agents across sentiment, prediction, risk, memory and execution over three trading venues
- Compressed a strategy question from an overnight job to 11.1s with a vectorized backtest replay
- Gated the live path behind win-rate thresholds, drawdown kill-switches and budget caps. Every position remains paper
- Replayed 89 × 250 × 24 to reach that figure, fast enough to stay inside a research loop
- Logged every agent decision to a pgvector telemetry store, feeding signal quality and win rates into tuning
- Built the lab across 59,393 lines and 668 tests, with the live execution path closed until a strategy clears
Professional Experience
Strategy & Operations Analyst, Accrete Consulting
- Supported an SAP ECC to S/4HANA conversion across 30+ legal entities, cutting month-end close from 10 days to 6
- Recovered the undocumented reporting logic behind 100K+ client records and rebuilt it in SQL, Python and Tableau
- Automated multi-step consolidation into dashboards, cutting reconciliation errors 30% across entities
- Configured and validated SAP FICO across AP, GL, intercompany and fixed assets, owning migration validation and UAT
- Surfaced anomalies the client had read as process error, and standardized OKR and KPI tracking across entities
GTM Campaign Manager, Analyst & Recording Engineer, NOGXNRE Records
- Drove full-cycle go-to-market for 5 artists, growing one release 100K to 1M in two weeks on playlist placement
- Engineered VBA royalty tracking across a $100K+/month catalog, recovering $1K-$10K in missed payments each cycle
- Grew catalog revenue 14% over one year on rolling streaming and sync payouts across the same catalog
- Ran paid activation across Spotify, YouTube and Instagram, coordinating release timing, creative and distribution
- Mixed and mastered for major-label artists including DIVINE, Karan Aujla, Jonita Gandhi, PapiTHBK, Samara Cyn and Prznt
Business Strategy & Operations Analyst, Influencer Management & Media Agency
- Delivered monthly executive reporting on OKR/KPI frameworks (MRR, LTV/CAC, churn) for 10+ creator-economy clients
- Standardized deliverables and launch-readiness tracking into templates, cutting reporting overhead 30%
Strategy & Data Analyst, Zara
- Replaced manual reporting for 300+ SKUs with a SQL-fed VBA dashboard, cutting the weekly cycle from 3 days to 1
Technical Skills
AI Engineering (built)
Agent Orchestration from Primitives · Multi-Agent System Design · Agentic Loops with Verification Gates · Hybrid Retrieval (Vector + AST Dependency Graph) · MCP Integration & Exposure · Function & Tool Calling · Context Engineering · Prompt Architecture · Structured Outputs · LLM Evaluation & Adversarial Testing · Requirements Traceability · Local Inference · FastAPI Services
Languages & Tools
Python · TypeScript / JavaScript · Rust · SQL · Node.js · React · Next.js · Tauri · FastAPI · asyncio · PostgreSQL · pgvector · SQLite · LanceDB · MLX · pytest · vitest · Playwright · Vite · Cloudflare Workers · Git · Tableau · Advanced Excel (VBA)
Platform & Delivery
Authentication & Authorization (OAuth 2.0, WebAuthn) · Access Control & Permissioning · Threat Modeling · Fail-Closed Design · Secrets Management · API Design · Data Modeling · CI Gating · WCAG 2.2 AA Accessibility · Zero to One Delivery
Familiar With
LangChain · LlamaIndex · LangGraph · CrewAI · Docker · agent observability and tracing tooling
Education & Certifications
Loyola Marymount University
B.A. Business Administration, Entrepreneurship
Certifications
AWS AI Practitioner (in progress, September 2026) · Bloomberg Market Concepts · DeepLearning.AI Generative UI Agents