Open to forward deployed engineering roles

I sit with the customer,
then ship what they run on.

Cole Bonawitz

I'm Cole, a forward deployed engineer. I scope workflows with the people who do them, then build the agents, data pipelines and integrations that take them over, with guardrails, evals and a human in the loop from day one.

See what I've shipped Résumé GitHubLinkedIn LinkedIn
~/deployments/coi-agent — multi-office insurance agency
running
  1. Inbound emailCOI requests
  2. Classifyopenclaw agent
  3. Match policypython mcp · 10 tools
  4. Risk + injectionrules in code
  5. Draft certificateacord · cover email
  6. Approvalfastapi dashboard
  7. Audit loghash-chained
Live diagram of the certificate-of-insurance agent I'm deploying now. Read the deployment →10 MCP tools43 testsdraft-only
Stack in production
ClaudeModel Context ProtocolPythonFastAPIPytestPostgreSQLSupabaseSQLiteReactTypeScriptDockerGitHub ActionsVercelGitHubFirebaseArduinoHubSpotZapier
01Output

What the work produced.

Every number comes from a deployment on this page. Click one to read it.
10-tool
MCP server in a live agent
↳ coi-agent
43 + 7
unit tests + golden evals
↳ coi-agent
20
agents on one memory layer
↳ cole-os
98%
accuracy, device deployed in 4 hospitals
↳ uganda
10,000+
records through a production pipeline
↳ swoop
$135K
revenue on a company I built
↳ rize
02Deployments · 08

Everything I've built and shipped.

Each one starts with who I sat with and what they needed, then what I built and what it did.
01AI Agents for Insurance Agencies·Oct 2026 — present
Deploying

Certificate-of-insurance agent

Founder & forward deployed engineer

An agent that reads inbound certificate requests, matches them to the right policy and drafts the certificate, so an agency's service team approves instead of retyping.

The customer

Scoped the certificate-of-insurance workflow with the VP of a multi-office independent agency: how requests arrive, which ones are risky, and what a person must always see before anything leaves.

What I built
  • +An OpenClaw agent and a 10-tool Python MCP server that classifies inbound email, matches the policy and drafts the certificate.
  • +Draft-only by design: every certificate stops in a FastAPI approval dashboard, with risk rules and a prompt-injection check written in code, not in the prompt.
  • +A hash-chained audit log of what came in, what was drafted and who approved it, plus 43 unit tests and a 7-case golden eval set.
10
MCP tools
43
unit tests
7
golden evals
0
sends without a human
Model Context ProtocolPythonFastAPIPytestClaudePostgreSQL
See the product page →
02Dignity Properties·Project
Shipped

Off-market deal sourcing agent

Builder

An agentic system for a real estate firm that automates off-market deal sourcing and remembers every property, owner and past decision.

The customer

A real estate firm wanted off-market deal sourcing automated, with the context about each property and owner kept between runs instead of rediscovered every time.

What I built
  • +Agents built with Claude Code and MCP that source and qualify off-market properties.
  • +A persistent memory layer in Supabase of properties, owners and past decisions, so each run starts from what the firm already knows.
MCP
tool layer
Supabase
persistent memory
ClaudeModel Context ProtocolSupabasePython
03Personal infrastructure·2026 — present
Running

Cole OS: personal agent operating system

Builder

The memory layer for a 20-agent fleet: one git vault every agent reads, with scopes, budgets and guardrails enforced by code.

The customer

My own problem, run like a customer's: dozens of scheduled agents, no shared memory, and no way to see or cap what each one could touch or spend.

What I built
  • +Tiered memory in a git vault, with per-agent scopes enforced by a pre-commit hook.
  • +Script-gated jobs with spend caps, so an agent cannot run or spend outside what it was given.
20
agents
per-agent
scopes in a hook
capped
spend per job
PythonSupabaseGitHubClaude
04DukeEngage Uganda·Jun — Aug 2025
Shipped

Hospital sensor device and data pipeline

Project manager

Embedded in four hospitals in Kampala to find the problem worth solving, then built and deployed a $25 device that saves nurses 2+ hours a day.

The customer

Interviewed 20+ clinicians across 4 hospitals and scored 40 pain points before writing any code, so we built for the problem nurses actually had.

What I built
  • +Benchmarked 3 sensor types and wrote the Arduino-to-Python ingestion pipeline.
  • +Built a 5,000-reading labeled dataset and deployed the device at 98% accuracy.
4
hospitals
40
pain points scored
98%
accuracy
2+ hrs
saved per nurse daily
ArduinoPythonpandas
05Swoop Tutors·Aug — Dec 2025
In production

First data pipeline and custom CRM

Go-to-market engineering intern · sole engineer

The first data pipeline for a 4-person startup, from raw records to a custom CRM the team still runs on.

The customer

Reported to the CEO twice a week. The team was tracking replies by hand and did not know which customers converted.

What I built
  • +Clay ingestion, SQL and pandas cleaning of 10,000+ records, ICP scoring that qualified 3,000 accounts, and HubSpot and Zapier sync.
  • +A Python + React reply-rate dashboard by segment that showed smaller consumer accounts converting best and changed who the team targeted.
  • +The relational schema and migration from HubSpot to a custom CRM that is still in use.
10,000+
records cleaned
3,000
accounts qualified
15% / 40%
cold / warm reply
50
calls booked
PythonpandasPostgreSQLReactHubSpotZapierClayClay
06RiZE Maps·Jun 2023 — Aug 2026
Shipped

Content engine and outbound agent

Founder

Built a DTC brand to $135K in revenue, then the systems that ran its growth: a content scoring engine and a draft-only B2B outbound agent.

The customer

Every system here had a real P&L behind it. I modeled landed COGS across 9 SKUs into a break-even ROAS rule per product line and lifted average order value from $55 to $90.

What I built
  • +A Python + SQLite engine that scores 135 reels against the account's own baseline, classifies hooks and drafts weekly ad briefs under hard Meta spend caps.
  • +A draft-only outbound agent (Python, SQLite, Claude) for 118 hotels in F1 host cities that verifies contact routes and gates every draft through 14 compliance rules and human approval.
$135K
lifetime revenue
3,500+
orders
3M+
organic impressions
14
compliance rules
PythonSQLiteClaudeMetaShopify
07Momnt·Oct 2025 — May 2026
Shipped

Usage instrumentation and retention loop

Co-founder & head of product

A digital-wellness NFC wearable and iOS app, taken from 300+ customer interviews to 300+ units sold.

The customer

Ran 300+ face-to-face discovery interviews to define the user, then converted a 200-user campus pilot into 300+ units and $9K+ in revenue. Pitched 5 VCs and won $7K across two competitions.

What I built
  • +Per-session usage instrumentation in Firestore.
  • +Python that auto-surveys users inactive for 7+ days. Fixes tested on a 50-user cohort drove 85% two-month and 70% six-month retention.
300+
interviews
300+
units sold
85%
2-month retention
70%
6-month retention
FirebasePython
08HackDuke · Finance track·24 hours
Shipped

Equity crowdfunding marketplace

1st place of 100+ teams

Designed the Postgres schema and the LLM API data pipeline for an equity crowdfunding marketplace, shipped in under 24 hours.

The customer

A hackathon brief where the scope had to fit a single day, so the data model had to be right the first time.

What I built
  • +The Postgres schema and the LLM API data pipeline behind the marketplace.
1st
of 100+ teams
< 24h
to ship
PostgreSQLPythonReact
03How I work

Every deployment runs the same loop.

Embed with the customer, ship behind guardrails, and let real corrections make it better.
  1. 01

    Embed

    Sit with the people who do the work: the VP, the service desk, the nurses. Watch the workflow before touching code.

  2. 02

    Scope

    Turn what I saw into a written spec: inputs, decisions, failure cases, and what a person must always approve.

  3. 03

    Integrate

    Connect to the systems they already run through APIs and MCP tools, not a new app they have to learn.

  4. 04

    Guard

    Risk rules, injection checks and approval gates in code. Agents draft; people decide.

  5. 05

    Evaluate

    Unit tests, golden eval sets and audit logs, so every change is measured before it ships.

  6. 06

    Iterate

    Real corrections become new eval cases and the system improves weekly with the customer.

04About

Engineer by training, founder by habit.

Cole Bonawitz speaking at an eventChattanooga, TN → Durham, NC

I'm a Computer Science student at Duke (minors in Statistical Science and Philosophy), studying abroad at UNSW Sydney this fall. Coursework in machine learning, data structures and algorithms, computer networks and probability.

The common thread in everything above: I go to where the work happens. Hospital wards in Kampala, a startup's sales floor, an insurance agency's service desk. I learn the workflow from the people in it, then build the system that takes the repetitive part off their hands, and I stay until it works.

Education
Duke CS · 3.95 GPA
Honors
Dean's List with Distinction
HackDuke
1st of 100+ teams
Founded
RiZE Maps · $135K
Agents & LLMs
Claude Code, MCP, OpenClaw, Hermes Agents, LLM APIs, evals, guardrails
Engineering & data
Python (pandas, FastAPI, pytest), SQL, Postgres, Supabase, React, TypeScript
Infra & integrations
Git, GitHub Actions, Docker, Vercel, REST APIs, Twilio, Clay, HubSpot, Zapier
Résumé (PDF) GitHubLinkedIn LinkedIn
Replies within 24h

Hiring forward deployed engineers?
Let's talk.

I want to be in the room with your customers, turning what they need into systems that ship. My résumé and the deployments above are the best summary of how I work.

cole.bonawitz@duke.edu LinkedIn LinkedInRésumé (PDF)