PD Product systems
for the agent era Start a conversation ↗

Punyaslok Dutta · Product engineer / agent operator

I turn customer mess into shipped product.

I stay on the problem, taste and tradeoffs. Agents fan out across research, architecture, implementation and QA—then evidence converges before I ship.

See selected work ↓ Watch the 60-sec version ↗
Operating belief

Parallelize discovery. Serialize decisions. Keep the customer in the loop.

CONTROL ROOM
ACTIVE BRIEF Reduce onboarding drop-off without adding a new step.
CONTEXT SCOUT

Turns call notes, support threads and product telemetry into a one-page evidence brief.

evidence.md
05sessions
03worktrees
01ship decision
LISTEN✳FRAME✳DIVERGE✳BUILD✳PROVE✳SHIP✳LEARN✳

01 / Selected product work

Three fintech chapters.
Products that had to work.

Consumer rewards, digital banking and AI-native compliance—shaped across product judgment, backend systems and delivery.

01 / May 2022—Dec 2023

Hubble Money

Rewards & payments app

Backend developer · 3-person backend team

Product surfaces shaped

Flexible
SIP flows Commerce
aggregator Razorpay
payments Analytics
activation
Customer problem → API design → launch → validation → scale
Scale context

10K → 500K+ customers in 6 months

Built for a product at 50K DAU and $5M AUM.
02 / Jan 2024—Aug 2025

barq

Digital wallet platform

Senior software engineer · Backend

Product surfaces shaped

Auth Customer
identity P2P
transfers Ops
tooling
Secure identity → money movement → recoverability → internal control
Operating context

20K DAU across critical financial services

Kafka retry design improved component-failure resilience by 7%; work supported ~10% monthly AUM growth.
03 / Aug 2025—Now

Settl X

AI-native AML platform

Co-founder & CTO · Product + engineering

Product surfaces shaped

BYOV
adapters Alert-triage
agent Compliance
memory SAR
drafting
Pilot signal → roadmap → product architecture → controlled deployment
Ownership

From customer workflow to production guardrails

Hybrid VPC, audit replay, citation provenance and SOC 2 Type 1 readiness.
One through-line: regulated money products where trust is a feature. Build · harden · learn · own more of the product

02 / The product loop

The job is not writing more code. It is making better decisions, faster.

Every phase leaves an artifact. That is how parallel work stays legible.

01

CUSTOMER SIGNAL → EVIDENCE

Start with what happened, not what was requested.

I review the call, pull the exact friction, and separate the user’s proposed solution from the job they need done.

Human owns
Curiosity, follow-ups, product judgment
Agent owns
Transcription synthesis, pattern scan, open questions
Exit artifact
evidence.md

03 / The control room

Many sessions.
One source of truth.

I use agents as a portfolio of bounded bets—not one endlessly growing chat.

00 Intake
SIGNAL“I keep abandoning setup.”Granola · support · analytics
01 Discover
SCOUT AMap the current flowread-only
SCOUT BFind prior patternsread-only
02 Build
TREE ASmallest pathfeature/a
TREE BRobust pathfeature/b
03 Prove
REVIEWERDiff against briefreview.md
QATry to break ittest.log
04 Decide
HUMAN GATEShip, revise or killdecision.md
01

One outcome per session

A thread gets a finish line, constraints and a clear return format.

02

Context packs, not context dumps

Each role sees only the brief, files and decisions it needs.

03

Branches do not talk. Outputs do.

Agents work independently; structured artifacts make their work comparable.

04

Merge through evidence

Tests, screenshots, traces and requirements—not confidence—win the decision.

04 / Attention allocation

Product and development are the work. GTM closes the loop.

I keep distribution intentionally small and high-signal: find the right customer, start a useful conversation, and route what I learn back into the product.

92%Product + development
8%GTM
PRODUCT / DEV

Customer problem · prototyping · architecture · coding · review · QA · release

GTM

ICP research · personal outreach · customer calls · X / LinkedIn notes

05 / The tool bench

A small core.
A deliberate bench.

Tools are hired for a job. Saved repos are research, not automatically the stack.

CORE

Granola

Customer calls become evidence, requirements and reusable language.

LISTEN / SYNTHESIZE
BENCH

Screenpipe

Local computer history as recallable context for long-running work.

MEMORY / CONTEXT
BENCH

Agent-Reach

Research across the public web and social surfaces from one agent path.

RESEARCH / SIGNAL
CORE

Cursor + coding agents

Fast local edits plus autonomous branches for bounded implementation work.

CODE / ITERATE
PATTERN

Git worktrees

Isolation for parallel builders without corrupting the shared working state.

BRANCH / CONVERGE
BENCH

qm · Tessera · Orca

Different control surfaces for multiplayer work, session organization and fleets.

ORCHESTRATE / WATCH
CORE

Tests + browser QA

Executable acceptance criteria, real flows and screenshots before merge.

VERIFY / SHIP
BENCH

OpenInference

Traces that show where an agent spent time, tokens and judgment.

OBSERVE / DEBUG
BENCH

SkillSpector

Security review for the capability packages that enter the system.

SCAN / TRUST
GTM

Clay

Build a small, qualified list around a real problem and a crisp ICP.

FIND / ENRICH
GTM

AgentMail + Mailometer

Personal outreach infrastructure with deliverability kept visible.

REACH / MEASURE
GTM

Granola → X / LinkedIn

Distill real customer language into an opinionated note, then shape it for each platform.

LEARN / TEACH
284public GitHub stars reviewed

Your saved work tells a useful story:

INTERFACESReact Native · audio · search→ INTELLIGENCEML · vision · generative systems→ LEVERAGEagents · skills · harnesses

06 / Design principles

The system gets faster.
The bar stays human.

01

Thin harness.
Fat skills.

Keep orchestration boring. Put method, judgment, examples and failure modes into reusable skills.

02

Deterministic floor.
Probabilistic ceiling.

APIs, scripts and tests handle certainty. Models interpret, compare and decide where judgment matters.

03

Disagreement
before code.

Two independent approaches are cheap before implementation and expensive after it.

04

Proof over
performance.

A confident agent is not evidence. A passing flow, a trace and a clean diff are.

05

Context is
a budget.

Load the right brief, decisions and files—not the entire history of the company.

06

Customer signal
should compound.

Every useful call improves the product, its examples, its skills and the next conversation.

I do not delegate

Problem definitionTasteTradeoffsCustomer trustFinal ship call

07 / Skills worth compounding

My real stack is a set of reusable ways of working.

  1. 01Customer-call synthesissignal → evidence brief
  2. 02Repository mappingcodebase → terrain map
  3. 03Architecture divergenceconstraints → options
  4. 04Implementation contractbrief → bounded task
  5. 05Adversarial code reviewdiff → ranked risk
  6. 06Release proofcriteria → receipts
  7. 07Learning captureincident → stronger skill
  8. 08Product story distillationtruth → demo / post

08 / The 60-second story

Show the operating system, not another talking head.

A sales video should move from customer signal to shipped proof in under a minute.

workflow-film_v01
00:06 · SIGNAL“I keep abandoning setup.”
00:18 · DIVERGE3 agents
2 approaches
1 brief
00:42 · PROOF8/8 flows passed
I use AI to compress the whole product loop—not just type faster.
00:0000:60
ProblemSystemProofInvite

09 / Build together

Have a hard product problem
and too many tabs?

punyaslok@settlx.xyz ↗

Tell me the customer, the constraint and what “working” needs to mean.