IRVING

c/o “software engineer”

Role: Senior RPA Developer
Company: Baker Tilly
Clearance: US Secret
Status: Available

Selected Work

[ 01 — 06 ]

AutoProspect

Draw a radius. Pick a category. A 4-phase Celery pipeline discovers, enriches 26 signals, and Claude-scores every lead. ~$0.02/business.

React 18ViteTailwind CSS v4Mapbox GL JSZustand +11

The Numbers

4 Pipeline Stages

4,647 Test LOC

~$0.02 Cost / Business

26+ Enrichment Signals

What I Built

Full-stack B2B SaaS for selling automation services to local businesses. Draw a radius on a Mapbox map, pick a category, and a 4-phase Celery pipeline fires: Google Places discovery → parallel website enrichment (26+ signals: CMS detection, CRM fingerprinting, booking tools, social scraping, review NLP) → Tier 1 Claude scoring → scan finalization. Celery chord orchestration fans out 50+ enrichment tasks in parallel then collapses into a nested scoring chord. Tier 1 scores every business automatically on four configurable dimensions (CRM, scheduling, marketing, invoicing) and returns structured JSON: 0–100 score, key signals, pitch angle, deal value estimate. Tier 2 triggers a deep-dive dossier on promoted leads — ROI argument, competitor analysis, implementation roadmap, objection rebuttals. Multi-tenant workspace isolation (8 Django apps), approval queues for AI-generated outreach, change detection that diffs enrichment snapshots across recurring scans, and per-call cost tracking that accumulates atomically on the parent scan.

Key Decisions

Celery chord orchestration fans 50+ enrichment tasks in parallel then collapses into a nested scoring chord; self-healing in score_business_tier1 detects dropped chord callbacks and finalizes the scan directly
Dynamic Claude prompts: scoring dimensions (4 weights summing to 1.0) and disqualification signals stored as JSON in AgentConfig, injected into the system prompt at runtime — zero code changes to customize scoring per workspace
3-day enrichment/scoring cache via enriched_at__gte cutoff — skips unchanged businesses on recurring scans, cutting Claude API spend ~2-3x; freshly enriched businesses always get a new score regardless
Atomic F() expressions on Scan.businesses_enriched prevent race conditions when 50+ concurrent tasks increment the progress counter simultaneously; circuit breaker auto-pauses SavedScans after 3 consecutive failures
Per-call cost tracking (prompt_tokens × model input rate + completion_tokens × output rate) saved to AutomationScore, then accumulated atomically on the parent Scan — budget enforcement before every new scan trigger

Where is my data?

181 datacenters. 43 cloud providers. Latency estimation. All on one map.

Next.js 16React 19TypeScriptMapbox GL JSZustand +2

The Numbers

181 Datacenters

43 Providers

33 Pricing Regions

What I Built

Interactive globe visualizing 181 cloud and colocation datacenters across AWS, Azure, GCP, Equinix, Digital Realty, Cloudflare, and 40+ other providers. Filter by provider type, estimate round-trip latency with a Haversine + fiber routing model, compare cloud pricing across 33 regions, and export filtered results as CSV or JSON.

Key Decisions

Zustand with 6 stores (datacenter, filter, map, comparison, latency, pricing) — no Redux boilerplate
Latency uses Haversine + 1.3x fiber overhead + hop decay + 15% congestion buffer — fully offline, no live ping
Uncontrolled Mapbox pattern — flyTo/setMaxBounds called imperatively via mapRef to avoid animation choppiness
Shareable URLs via custom useUrlSync hook with 500ms debounce — encodes filters, viewport, and selected datacenter into query params

Weather Arbitrage Dashboard

Real-time P&L, live log streaming, and 34 analytics endpoints for the trading daemon.

React 19TypeScriptVite 6Tailwind CSS 4Zustand +5

The Numbers

34 Analytics Endpoints

WS Live Streaming

10s Frontend Poll

What I Built

Full-stack monitoring dashboard for the Kalshi weather arbitrage daemon. Reads structured JSON/JSONL files written by the daemon, streams live logs over WebSockets, and renders sortable position tables, P&L charts, and backtest analysis. 34 analytics endpoints cover win rates, edge calibration, provider accuracy, and cost efficiency.

Key Decisions

File-reader architecture — no database; Django reads daemon-produced JSON/JSONL on every request for always-fresh data
Django Channels + Daphne for ASGI — WebSockets handle live log streaming, HTTP handles all structured data
Zustand (1KB) over Redux — 2 stores, full TypeScript inference, zero boilerplate
Exponential backoff WebSocket reconnect (1s → 2s → 4s → max 30s) on the React client

Weather Arbitrage Daemon

Five ensemble models. Bayesian blending. Quarter-Kelly sizing. Running 24/7.

PythonKalshi API v2 (RSA-PSS)NOAAOpen-Meteo (GFS/ICON/ECMWF/GEM)Bayesian Probability +2

The Numbers

5 Ensemble Models

11 Cities Covered

15m Poll Interval

20 Max Positions

What I Built

Autonomous Python daemon that identifies and trades mispriced temperature contracts across 11 US cities on the Kalshi prediction market. Blends forecasts from NOAA, GFS, ICON/DWD, ECMWF, and GEM using Bayesian log-odds at 30/70 model/market weighting, sizes positions with Quarter-Kelly criterion, and enforces circuit breakers at daily and weekly loss limits.

Key Decisions

Bayesian log-odds blending at 30% model / 70% market weighting — inverse-transformed back to probability space per trade
Quarter-Kelly criterion for position sizing from live edge and current bankroll — not a fixed fraction
RSA-PSS SHA256 signing for all Kalshi API calls using PEM private key — not API-key-only auth
Geography-based correlation groups (Gulf South, Northeast, Pacific) cap correlated exposure
Circuit breakers halt all trading at $5 daily / $10 weekly loss — daemon stops, does not just skip individual trades

Crypto Trading Bot

Five signals. Composite scoring. Ten pairs. Limit buys. Zero babysitting.

PythonCoinbase Advanced Trade APIRSI-14MACDEMA-12/26 +1

The Numbers

10 Trading Pairs

5 Composite Signals

30s Polling Interval

$50 Max Position

What I Built

Automated cryptocurrency trading daemon for Coinbase Advanced Trade API. Scores each of 10 pairs on a 0–100 composite signal scale using RSI-14, MACD, EMA-12/26, volume spikes, and support-bounce analysis. Places limit buys 0.1% below market and market sells for stop-loss exits. 30-second polling loop with Telegram alerts and JSON portfolio state.

Key Decisions

Composite 0–100 scoring gate — buys only trigger at ≥50 points to filter weak signals
All indicators (RSI-14, EMA-12/26, MACD) implemented from scratch in pure Python — zero pandas or numpy
Limit buys placed 0.1% below market for maker-fee advantage; market sells used only for stop-loss to guarantee fills
Max 5 concurrent positions, 30-minute per-pair cooldown, 40% balance cap per trade enforced before any buy

File-to-API Platform

Drop a CSV. Get a REST API with auth, pagination, and Swagger docs. Done in seconds.

Django 4.2Django REST FrameworkReact 19TypeScriptVite 7 +3

The Numbers

<5s Upload → Live API

1k/hr Rate Limit (reads)

10 MB Max Upload

What I Built

Upload a CSV or Excel file and get a fully functional REST API instantly. Django parses the file with pandas, infers column types at a 90% match threshold, creates a SQLite table dynamically at runtime using Python's type() constructor, and registers DRF ViewSets and URL routes without a server restart. Full CRUD, API key auth, rate limiting, and OpenAPI/Swagger docs included.

Key Decisions

Dynamic model generation at runtime via Python's type() constructor — no code generation, no server restart on upload
Runtime DRF router registration — ViewSets and URL routes injected into live URL conf without touching urls.py
SHA-256 + salt API key hashing with 8-char plaintext prefix for O(1) lookup — format: fta_{prefix}_{secret}
90% type-match threshold for schema inference — columns only typed as int/float/date if 90%+ of non-null values match
python-magic for MIME type validation — blocks malicious uploads regardless of file extension

Experience

[ 2020 — Present ]

ACTIVE

Aug 2025 – Present

Baker Tilly Senior RPA Developer & Automation Consultant

[ SECRET ]

Modernizing how the Navy processes 2 million fitness reports a year.

Leading a 3-dev team building end-to-end UiPath automation (Studio, Orchestrator, Document Understanding) for E1–E4 sailor fitness evaluation processing
Training custom ML models in UiPath AI Center for high-accuracy extraction from diverse PDF formats and legacy Navy docs
Automation projected to save the Navy millions annually in personnel costs
Working directly with Navy stakeholders inside Secret-level security requirements

COMPLETED

Jan 2024 – Aug 2025 · 1yr 8mo

ONeil Interactive Senior Backend Engineer

Backend infrastructure for high-traffic platforms serving the residential homebuilding industry.

Built and maintained scalable backend systems in Python/Django/DRF for homebuilder marketing platforms
Led development of Homefiniti CMS — clients' primary tool for managing property data at scale
Modernized legacy MLS data integrations, replacing brittle code with clean, maintainable patterns
Upgraded legacy Python/Django codebases to current versions, improving security and performance

COMPLETED

Jan 2022 – Jan 2024 · 2yrs

Lightmatter Software Engineer

Full-stack engineering in healthtech — HIPAA-compliant features shipped in a fast Agile environment.

Built a voice-to-text AI pipeline: OpenAI Whisper → fine-tuned BART classifier → GPT-3.5-Turbo structured output
Architected patient email automation using live EHR/EMR data for targeted, automated outreach
Led React Native and HTMX frontend work alongside Django/DRF/FastAPI backend development
Delivered across GCP, Heroku, Docker, Postgres stack with full ownership end-to-end

COMPLETED

Aug 2020 – Jan 2022 · 1yr 7mo

C Spire Jr. Software Developer

Started here. Shipped a core internal platform and migrated Perl to Django before the end of my first year.

Built Django platform for phone call management — query, share, and modify telecom rate center inventory reports
Led end-to-end development of customer data usage tracking app for billing cycle support
Migrated legacy Perl tools to Django with redesigned frontends — cleaner code, better UX
Operated in full Agile lifecycle: standups, sprint planning, retros

Capabilities

[ 004 ]

RPA & Automation

Enterprise-grade automation at Navy scale.

UiPath Studio Orchestrator Document Understanding AI Center Attended & Unattended Bots OCR & ML model training Python automation Selenium asyncio

Backend Engineering

APIs that don't fall over.

Python Django DRF FastAPI REST WebSockets PostgreSQL SQLite Docker GCP Heroku Linux systemd

Frontend & Full-Stack

I build the whole thing.

React 19 TypeScript JavaScript React Native HTMX Tailwind CSS Astro

Algo Trading & Systems

Running in production. With real money.

Algorithmic trading Kalshi API Coinbase Advanced Trade JSONL logging Telegram bots Circuit breakers Risk management asyncio event loops

Currently at Baker Tilly, always shipping something on the side.