Lead AI engineer · Agent harnesses · Cebu, Philippines

I turn ambitious AI ideas into reliable products that ship.

I scope, build, evaluate, and productionize agentic systems, RAG products, and AI-native workflows through Chalk—personally owning the hard parts between prototype and production.

Reserve the Blueprint See client outcomes

$500 · 90 minutes · Fully credited toward a build started within 14 days.

Personally ledScopeBuildEvaluateShip

David Brent PanonceLead AI engineer · Chalk

Operational proof

Receipts before résumé.

Open the case files

Result 0144K+production sessionsAgentic real-estate journeys handled in production.nōna production data

Result 0276%visit → active searchConversion into the core product journey.nōna product analytics

Result 037enforceable agent gatesExternal checks that refuse unsafe progress.Chalk Protocol

01The first engagement

01 / Start here

AI Build Blueprint

A focused working session that turns an AI opportunity, fragile prototype, or painful workflow into a production-ready plan you can act on.

Decision session

$50090 minutes · remote

Fully credited toward a build started within 14 days.

Reserve the Blueprint

One focused decision · no sales theatre

Outputs

What leaves the room with you

4 concrete artifacts

01
Opportunity and workflow map
02
Architecture and risk decisions
03
Prioritized production plan
04
Scope, estimate, and next-step recommendation

The outcome is a decision-ready production path—not another strategy deck.

A strong fit when

You have a real workflow, audience, or prototype to improve.
You need senior technical judgment before committing to a build.
Reliability, evals, and production constraints matter.

Probably not yet when

You only need generic AI training or brainstorming.
There is no clear user, workflow, or business owner yet.
You want a speculative demo with no path to production.

After the Blueprint

Production AI Sprint

From $5,000

A scoped build that takes the highest-value slice from working prototype to evaluated, production-ready release.

02Selected outcomes

Production evidence, not capability claims.

Three different constraints—production scale, trustworthy autonomy, and a hard launch window—with the same job: turn ambiguity into something people can rely on.

Production AI · PropTech

nōna

A multi-agent real-estate concierge built for real customer journeys—not a scripted demo.

Problem: Philippine property discovery is fragmented across listings, verification, scheduling, and owner workflows. The product needed to hold context across a non-linear journey without sacrificing speed or trust.
What I changed: I architected an eight-agent system with Google ADK and Gemini, a FastAPI coordination layer, and production state management designed around search, scheduling, verification, and recovery paths.
Result: The experience moved visitors into active property search while supporting production traffic without major outages during the measured launch period.

Visit nōna Read the case study

nōna / agent systemEvidence

SpecialistSearch

SpecialistVerify

SpecialistSchedule

SpecialistOwner

CoordinatorHome concierge

Shared context · routing · recovery

SpecialistListings

SpecialistFAQ

SpecialistLeads

SpecialistInsights

Eight specialist agents coordinate discovery, verification, scheduling, and owner workflows through a shared production state.

44K+production sessionsProduction data

76%visit → active searchProduct analytics

0major outagesOperations record

Google ADKGeminiFastAPINext.jsSupabaseCloud Run

Agent infrastructure · Open source

Chalk Protocol

A director’s harness that makes coding agents prove the work before they advance.

Problem: Coding agents can weaken visible tests, silently make product decisions, and still declare themselves done. Teams need a control layer outside the model that preserves intent and makes progress depend on evidence.
What I changed: At Chalk Agents, I led and built the durable .chalk project spine and zero-dependency CLI: locked acceptance tests, external verification, adversarial review, held-out audits, and human decision gates. I also built Chalk Code as the private R&D layer for model routing, compatibility, evidence, and containment across coding harnesses.
Result: Chalk Protocol now drives a provider-neutral development lifecycle across four first-party agent integrations, refuses unsafe progress at the command boundary, and dogfoods the same gates on its own development.

Explore Chalk Try the Protocol View the source

Chalk / trust architectureEvidence

Director intentExternal proof required

01Read

02Work

03Verify

04Write

P1 Criteria

P2 Spec

P3 Break-it

P4 Verify

P5 Review

P6 Integrity

P7 Audit

Public · OSSchalk/0

Chalk Protocol

Durable project state, command gates, independent review, and held-out regression.

Claude · OpenCode · Codex · Gemini

Private R&D

Chalk Code

Model routing, compatibility guards, tool evidence, benchmark probes, and fail-closed containment.

Route · Observe · Contain · Prove

The public protocol supplies durable state and hard gates; Chalk Code explores routing, evidence, compatibility, and containment beneath the harness.

7enforceable gatesProtocol specification

4first-party agent CLIsProvider matrix

0runtime dependenciesPublished package

Node.jsAgent adaptersLocked testsAdversarial reviewHeld-out evals

Rapid build · Civic tech

Senator Match

A neutral voter-matching product shipped inside a one-week election window.

Problem: Voters needed a fast way to compare senatorial candidates against their priorities, with very little time left before the 2025 Philippine election.
What I changed: I built the questionnaire, RAG-assisted matching flow, result generation, and production deployment with a deliberate focus on speed, neutrality, and shareability.
Result: The product launched four days before the election and reached more than one thousand users during its first day live.

Open the project

Senator Match homepage inviting voters to find candidate matches — Public product experience with launch analytics used as a supporting evidence inset.

Analytics showing Senator Match traffic increasing after launch — Public product experience with launch analytics used as a supporting evidence inset.

1K+users in 24 hoursGoogle Analytics

1 weekbuild to launchProject timeline

Next.jsFastAPIMongoDBOpenAILangChainAWS Lambda

03How the work moves

A production path with decisions at every handoff.

Each stage ends in something inspectable, so momentum never depends on a demo looking convincing.

0101 / Direction

Scope

Opportunity brief

Pin down the user, workflow, value, constraints, and smallest production-worthy slice.

0202 / Execution

Build

Working product slice

Ship the real workflow with deliberate architecture, not a disposable happy-path demo.

0303 / Execution

Evaluate

Eval and risk report

Test quality, failure modes, latency, cost, and recovery before customers discover them.

0404 / Release

Ship

Production handoff

Deploy with observability, operating notes, ownership, and the next iteration already clear.

04The operator

AI velocity, with an engineer still holding the architecture.

Senior product judgment stays attached to the implementation—from the first architecture decision to the production handoff.

I came up through mobile banking, production migrations, and CTO roles before specializing in agentic systems. That background shapes how I build AI: move fast, but keep reliability, security, and maintainability visible.

I work from Cebu with clients across time zones, co-lead the AI Cebu community, and teach the same multi-agent and production patterns I use in shipped systems.

Read the full résumé

Working principle

Move fast, while keeping reliability, security, and maintainability visible.

years shipping production software

20+

mobile builds shipped in two months

technical articles on production agents

AI Cebu

community co-lead and speaker

Selected talk · July 02, 2026Building Multi-Agent Systems Selected talk · March 28, 2026Skill Issue: Level Up Your AI Agents

05Before we start

Clear answers before a call.

The Blueprint should remove ambiguity—not create another sales process. These are the decisions most teams need before they can move.

What happens in the AI Build Blueprint?

We work through your users, workflow, data, risks, architecture, and fastest production-worthy slice. You leave with a concrete opportunity map, technical direction, prioritized plan, and estimate—not a generic strategy deck.

Is the $500 fee credited toward implementation?

Yes. If we begin an agreed Production AI Sprint within 14 days, the full Blueprint fee is credited toward that build.

Can you work with an existing prototype or team?

Yes. I can assess a fragile prototype, own a high-risk product slice, or work alongside an existing engineering and product team. The Blueprint is used to define boundaries and ownership before work starts.

Who owns the code and outputs?

Client work is structured so the agreed code and project artifacts transfer to the client. Any reusable Chalk tooling or pre-existing components are identified explicitly before the engagement.

Do you only use one model or framework?

No. I choose models and infrastructure around the workflow, quality bar, latency, data, and operating constraints. Google ADK, Gemini, OpenAI, Claude, FastAPI, Next.js, and Supabase are tools—not the offer.

What if the Blueprint shows AI is the wrong solution?

That is a useful outcome. You still receive the workflow and architecture analysis, plus a recommendation for the simplest credible path—even when that path uses less AI than expected.

How do timezone and communication work?

I am based in Cebu at UTC+8 and work asynchronously with scheduled overlap for clients in Asia, Europe, and North America. Response expectations and meeting windows are agreed before a build begins.

One clear first step

Turn the opportunity into a plan you can ship.

AI Build Blueprint · $500 · 90 minutes. Leave with the workflow, architecture, risks, and next production move made explicit.