Product Engineer · Creative Technologist · AI Product Builder

Ashish Gupta

Building AI-native products people actually want to use.

From conversational AI and intelligent analytics to realtime systems and developer platforms — I design products where the engineering disappears behind exceptional experiences.

Currently
Senior Product Engineer, AI — Shiprocket
Track record
7 years · 20+ products shipped
Open to
Founding & AI product engineering roles

Selected Work · 2019 — 2026

Products, not projects.

Eight builds that went from a blank document to production — each one researched, designed, argued over, engineered, and shipped.

MeetAira — character platform2024 — 25

01

MeetAira

A companion that listens, remembers, and speaks — in real time.

An AI emotional companion and character platform. Custom speech-to-text and text-to-speech pipelines, Live2D-driven expression, and a streaming architecture engineered so a spoken reply is never more than a breath away.

  • Voice AI
  • Custom TTS · STT
  • Live2D
  • Streaming
  • FastAPI

Shiprocket Trends — analytics platform2022

02

Shiprocket Trends

Analytics that answers questions, not just renders charts.

An enterprise analytics platform for e-commerce sellers — AI-generated insights over interactive dashboards at SKU and city granularity, tuned until the heaviest query still felt instant. It replaced a culture of exported spreadsheets.

  • AI Insights
  • Data Viz
  • Enterprise UX
  • Performance

Shiprocket Copilot — AI assistant2023 — 24

03

Shiprocket Copilot

An assistant that does the work, not just answers questions.

A realtime voice and chat copilot embedded in the seller back-office. It retrieves knowledge, streams answers, and executes tools against live order workflows over WebRTC — currently handling 10,000+ queries a month.

  • Realtime Voice
  • Tool Execution
  • WebRTC
  • RAG
  • Streaming

Dockyard — internal AI platform2024 — 25

04

Dockyard

The platform every AI company builds eventually — built deliberately.

An internal AI platform: LLM orchestration behind one gateway, prompt management with versioning, per-team budgets and governance, a model playground, RAG pipelines, and end-to-end observability for every token spent.

  • LLM Orchestration
  • Prompt Mgmt
  • Governance
  • RAG
  • Observability

TwentyTwo — AI phone agent2025 — 26

05

TwentyTwo

A phone agent that recovers revenue while the store sleeps.

An AI phone agent for Shopify merchants. It calls abandoned carts, holds a natural voice conversation inside merchant-defined guardrails, applies recovery offers, and writes every outcome back into the store's workflow.

  • Voice Workflows
  • Shopify
  • Automation
  • Telephony

Agentic Component Library — design system2023 — 25

06

Agentic Component Library

One design system, two frameworks, zero drift.

A design system shipped to React and Angular from a single source of truth — design tokens, accessibility contracts and performance budgets built into every component, with migration tooling so teams adopt it without stopping.

  • Design System
  • React
  • Angular
  • A11y
  • Tooling

Angular Migration Framework2024

07

Angular Migration Framework

Migrating a decade of enterprise UI without stopping the ship.

A framework for migrating a large enterprise Angular codebase incrementally: an extraction graph over the legacy code, automated component mapping, a verification pipeline, and a parity system that proves each migrated screen behaves identically.

  • AST Tooling
  • Codemods
  • Verification
  • Parity System

Live Orders Dashboard — realtime leaderboard2023

08

Live Orders Dashboard

Enterprise data with race-day energy.

A realtime order leaderboard inspired by MotoGP timing screens — live rankings and position-change choreography across thousands of updates a minute, without dropping a frame or losing legibility.

  • Realtime
  • WebSockets
  • FLIP Animation
  • Enterprise Dashboard

Engineering Philosophy

Six things I refuse to compromise on.

01 / 06

Beautiful products are engineered.

The polish you can see is a byproduct of decisions you can't — rendering budgets, data models, failure states, latency contracts.

02 / 06

Great UX begins long before pixels.

In the problem statement, the data model, the latency budget. By the time a design file is opened, most of the experience is already decided.

03 / 06

Performance is a feature.

Users don't separate speed from quality. A hundred milliseconds is the difference between a tool and a toy.

04 / 06

Animation must communicate.

Motion is information — what changed, where it went, what matters now. If it can't explain itself, it's decoration. Cut it.

05 / 06

Complexity should disappear.

The hardest engineering is invisible — streaming, retries, reconciliation, recovery. What users receive is a product that simply feels certain.

06 / 06

Users should never notice the engineering.

They should only notice that they finished.

Experience · Seven Years

From first commit to founding builds.

  1. 2019

    First production code

    Frontend engineering for e-commerce at scale. The obsession with craft starts here.

  2. 2020

    Owning outcomes

    From tickets to products — shipping complete features end to end, measured on user impact.

  3. 2021

    Design systems

    Building the component infrastructure other teams ship on. Accessibility and performance become defaults.

  4. 2022

    Shiprocket Trends

    First zero-to-one platform: self-serve analytics for one of India's largest e-commerce ecosystems.

  5. 2023

    Voice AI

    A realtime voice hackathon build places top-5 (Azure OpenAI) — and becomes the seed of Shiprocket Copilot.

  6. 2024

    Dockyard

    Internal AI platform: orchestration, prompt management, budgets, governance, observability.

  7. 2025

    MeetAira

    Founding build — a realtime AI companion with a custom voice stack and character platform.

  8. 2026

    TwentyTwo

    AI phone agents that recover revenue for Shopify merchants. Zero to one, again.

7
Years building
20+
Products shipped
4
Platforms from zero
2
Founding AI builds

Shiprocket

Senior Product Engineer — AI Platforms

2022 — Present

  • Took Trends, Copilot, Dockyard and the Live Orders Dashboard from concept to production across analytics, assistants and internal AI infrastructure.
  • Led AI platform engineering: LLM orchestration, retrieval pipelines, voice, evals and cost governance for product teams across the company.
  • Built the cross-framework design system (React + Angular) and the migration tooling that moved legacy enterprise UI onto it.

Independent products

Founder-engineer — MeetAira, TwentyTwo

2024 — Present

  • MeetAira: an AI emotional companion — custom TTS/STT, Live2D expression, streaming conversation architecture.
  • TwentyTwo: AI phone agents for Shopify cart recovery — telephony, voice workflows, merchant automation.
  • Everything from positioning and pricing to infrastructure and observability. Zero to production, alone or with tiny teams.

Earlier

Product & frontend engineering — e-commerce and SaaS

2019 — 2022

  • High-traffic consumer frontends: performance engineering, accessibility, and the discipline of shipping to millions.
  • Grew from executing specs to writing them — and learned that the spec is the product.

About

Engineer by craft, product person by conviction.

I'm Ashish — a product engineer who believes the distance between a good idea and a great product is engineering taste: knowing what to build, what to refuse, and what to make invisible.

For seven years I've worked across the full surface of product — strategy, UX, frontend architecture, distributed backends, and the AI infrastructure underneath. Not because generalism is fashionable, but because the products I want to exist can't be built from inside one lane.

Right now that means AI-native products: voice companions, copilots that execute real work, platforms that make LLMs safe and affordable at company scale. The common thread — people should feel the product, never the machinery.

Portrait — B&W editorial · placeholder

Product & Interface

The last ten percent that takes ninety percent of the time — that's the part users touch.

  • React
  • Next.js
  • TypeScript
  • Motion
  • Accessibility
  • Performance

Systems & Realtime

Streaming-first backends where milliseconds are the budget.

  • FastAPI
  • NestJS
  • Node
  • Redis
  • Postgres
  • WebRTC
  • WebSockets

Applied AI

From prompt to production — and the unglamorous plumbing in between.

  • LLMs
  • RAG
  • Voice AI
  • Streaming
  • Prompt Systems
  • Evals
  • OpenAI

Developer Experience

Design systems and platforms that make other engineers faster than me.

  • Design Systems
  • Architecture
  • Migration Tooling
  • Observability
  • CI Verification

Lab

Experiments worth losing sleep over.

Some of these become products. Some become lessons. Both are the point.

  • L·01 Sub-second voice turn-taking Interruption-aware duplex audio — the agent stops talking the instant you start. Exploring
  • L·02 Live2D avatar rendering pipeline Expression-mapped characters driven by conversation sentiment, at 60fps in the browser. Shipped · MeetAira
  • L·03 Prompt-diff CI Regression tests for LLM behaviour — every prompt change runs against a golden conversation set before merge. Prototype
  • L·04 Generative UI Interface components streamed from model output — layout as a first-class model capability. Exploring
  • L·05 Race-timing physics for data The MotoGP leaderboard choreography, generalised into a reusable realtime-ranking motion system. Shipped · Live Orders
  • L·06 On-device STT fallback Keeping voice products conversational when the network isn't — local-first speech recognition with seamless handoff. Exploring

Contact

Let's build something extraordinary.

The most interesting products start as slightly uncomfortable ideas. If you're sitting on one — as a founder, an investor, or a team that ships — I'd like to hear it.

Prefer a conversation? Book a call — twenty minutes, no slides.

Open to
Founding Engineer
And
AI Product Engineering
And
Technical Leadership
Selectively
Consulting
01 · MeetAira

A companion that listens, remembers, and speaks.

MeetAira is an AI emotional companion — a character platform where conversations happen in real time, by voice, with a face that reacts while you speak.

Role
Founding build — product, platform, AI
Timeline
2024 — 2025
Stack
FastAPI · WebSockets · Custom TTS/STT · Live2D · Redis
Status
Live

01Problem

Companion apps promise presence and deliver a chat box. Text is too slow for intimacy, and off-the-shelf voice stacks add three to five seconds of dead air per turn — which reads as indifference. The product question wasn't "can an LLM chat" but "can a machine hold a conversation that feels alive."

02Discovery

Prototyping showed the emotional ceiling was set almost entirely by latency and voice quality — not model intelligence. Users forgave a shallow answer delivered instantly with warmth; they abandoned brilliant answers that arrived late in a flat voice. That inverted the roadmap: the voice pipeline became the product, the model became a component.

03Architecture

CLIENT MIC · LIVE2D GATEWAY WS · DUPLEX STT PARTIALS DIALOGUE LLM · MEMORY SEGMENTER SENTENCES TTS CHUNKED AUDIO EXPRESSION EVENTS

04Engineering

  • Speak-while-thinking. The LLM streams tokens into a sentence segmenter; each sentence is synthesized and played while the next is still being generated. First audio lands before the model finishes its thought.
  • Custom TTS/STT. Off-the-shelf voices broke the character illusion. A tuned voice stack gave each character a consistent identity — and cut per-minute cost enough to make long sessions viable.
  • Live2D expression channel. Sentiment and phoneme events ride a side-channel to the renderer, so the face reacts during speech, not after it.
  • Interruption handling. Barge-in detection cancels synthesis mid-sentence and re-plans the turn — the difference between a recording and a conversation.
# the core loop: audio out begins before the model finishes
async def speak(turn):
    async for text in llm.stream(turn.context):
        for sentence in segmenter.push(text):
            audio = tts.synthesize(sentence)   # ~120ms to first chunk
            await ws.send(audio)               # play while thinking
            emit_expression(sentence.sentiment)

Streaming turn pipeline — simplified

05Impact

<1s
To first spoken audio
60fps
Live2D expression in-browser
Full stack
Custom voice pipeline, owned end to end

06Lessons

Emotional products are latency products. Every architectural decision — segmentation, chunked synthesis, barge-in — existed to protect a feeling, not a metric. When the engineering worked, nobody mentioned the engineering. That was the goal.

02 · Shiprocket Trends

A self-serve analytics platform for e-commerce sellers — interactive dashboards at SKU and city granularity, with an AI insight layer that says what the charts mean.

Role
Product engineer — end to end
Timeline
2022, evolved since
Stack
React · TypeScript · NestJS · Charting · LLM insights
Status
Live in production

01Problem

Sellers were exporting spreadsheets to answer basic questions: which SKU is growing, which city is slipping, what changed this week. Analysts became a bottleneck; decisions ran days behind the data. The existing tools rendered charts — they didn't answer questions.

02Design

The organizing principle: every screen must answer a question a seller actually asks, in their words, in under three seconds. Dashboards were designed backwards from those questions. Filters compose instead of stacking; drill-downs preserve context; and an insight strip narrates each view — "West-zone prepaid orders up 23%, driven by two SKUs" — so the reading is done for you.

03Architecture

EVENTS ORDERS · SHIPMENTS AGGREGATION SKU × CITY × DAY QUERY LAYER CACHED · <300MS DASHBOARD REACT · VIRTUALIZED INSIGHT ENGINE LLM NARRATION

04Engineering

  • Pre-aggregation over cleverness. The heaviest questions were known in advance, so the pipeline materializes SKU × city × day rollups — the UI never waits on a raw scan.
  • Perceived performance. Skeleton-free rendering: charts draw from cache immediately, then reconcile. No spinners on the money screens.
  • Virtualized everything. Ten-thousand-row tables filter and sort at 60fps because only what's visible exists.
  • Insight guardrails. The LLM narrates only computed aggregates — it never invents numbers; every claim links to the chart that proves it.

05Impact

~80%
Less manual spreadsheet work
SKU × City
Granularity, self-serve
<3s
Question to answer, p95

06Lessons

Enterprise users don't want more data — they want fewer decisions. The AI layer succeeded precisely because it was constrained: narration, never invention. Trust, once earned, made everything else adopted.

03 · Shiprocket Copilot

An assistant that does the work.

A realtime voice and chat copilot embedded in the seller back-office — it retrieves knowledge, streams answers, and executes tools against live order workflows.

Role
Lead engineer — AI & product
Timeline
2023 — 2024
Stack
Next.js · NestJS · OpenAI · WebRTC · MCP · RAG
Status
Live · 10K+ queries/month

01Problem

Sellers lived in support queues for actions the system could perform in seconds — "where is order 4127", "reschedule this pickup", "why was my COD remittance short". A chatbot that answered questions would have been a FAQ with better manners. The bar was an assistant that finishes the task.

02Discovery

Shadowing support conversations produced a taxonomy: roughly seventy percent of queries mapped to twenty tool-shaped actions; the rest needed knowledge retrieval. That split defined the architecture — a tool-execution loop for actions, RAG over help content and seller data for the rest, and one conversational surface over both. A hackathon voice prototype (top-5, Azure OpenAI) proved the voice channel deserved to be first-class, not a gimmick.

03Architecture

CHAT UI STREAMED VOICE WEBRTC RT EDGE WS · SFU ORCHESTRATOR LLM TOOL LOOP ORDER APIS GUARDED WRITES KNOWLEDGE RAG INDEX AUDIT LOG EVERY TOOL CALL
// the loop that turns answers into actions
while (!turn.done) {
  const step = await model.stream({ messages, tools });
  if (step.toolCall) {
    // risky writes (refunds, cancellations) require confirm-first
    const result = await execute(step.toolCall, { confirm: isRisky(step) });
    messages.push(result);
  } else turn.done = true;
}

Tool-execution loop — simplified

04Engineering

  • Confirm-first writes. Reads execute silently; mutations render a preview card the seller approves. Trust was designed, not assumed.
  • One pipeline, two channels. Voice and chat share the orchestrator — voice adds STT partials in and sentence-chunked TTS out, nothing forks.
  • Retrieval with receipts. Every knowledge answer cites its source document inline; unverifiable answers are refused by policy.
  • Full audit trail. Every tool call is logged with inputs, outputs and the conversation that authorized it.

05Impact

10K+
Queries handled monthly
70%
Of intents resolved by tools, not text
2
Channels, one orchestration pipeline

06Lessons

The hard part of an agent isn't the loop — it's the social contract around writes. Preview cards and audit logs did more for adoption than any model upgrade. People don't adopt intelligence; they adopt accountability.

04 · Dockyard

One gateway for every model.

Dockyard is an internal AI platform — orchestration, prompt management, budgets, governance, a playground and observability — so product teams ship AI features without each rebuilding the plumbing.

Role
Platform architect & lead engineer
Timeline
2024 — 2025
Stack
FastAPI · Postgres · Redis · Multi-provider LLMs · OTel
Status
Internal production

01Problem

Five teams, five OpenAI keys, five retry implementations, zero visibility. Prompts lived in code, spend lived in surprise invoices, and every incident began with "which team's integration is this?" AI adoption was outpacing AI governance by quarters.

02Design

The platform bet: make the governed path the easiest path. One SDK call gets routing, retries, fallbacks, caching, budget enforcement and tracing — doing it yourself becomes the expensive option. Prompts are versioned artifacts with owners and rollout stages, not string literals. The playground runs against the same gateway as production, so what you test is what ships.

03Architecture

TEAM APPS ONE SDK GATEWAY AUTH · BUDGETS ROUTER COST · FALLBACK PROVIDER A · B · C HOSTED + SELF-HOSTED RAG SERVICE SHARED INDEXES PROMPT REGISTRY VERSIONED · STAGED TRACES OTEL · EVALS

04Engineering

  • Budgets that degrade gracefully. Teams hitting limits fall back to cheaper models with a warning — features slow down before they black out.
  • Prompt lifecycle. Draft → staged → production, with diffs, owners and instant rollback. A prompt change is a deploy, and it's treated like one.
  • Trace-per-request. Every call carries prompt version, route, tokens, latency and cost through OpenTelemetry — incidents start with evidence, not archaeology.
  • Provider abstraction that leaks on purpose. Teams can pin models when they need to; the platform routes when they don't care. Escape hatches kept adoption voluntary — and total.

05Impact

1
Gateway for all AI traffic
Per-team
Budgets, enforced at the edge
100%
Of calls traced, prompt-versioned

06Lessons

Internal platforms live or die on the default path. Nobody was forced onto Dockyard; it won because the governed way was also the lazy way. That's the whole trick, and it's harder than it sounds.

05 · TwentyTwo

Revenue recovery, by phone.

TwentyTwo is an AI phone agent for Shopify merchants — it calls abandoned carts, holds a natural conversation within merchant guardrails, and writes every outcome back to the store.

Role
Founder-engineer
Timeline
2025 — 2026
Stack
Voice pipeline · Telephony · Shopify APIs · Workflow engine
Status
In build · early merchants

01Problem

Roughly seven of ten carts are abandoned, and merchants fight it with the two channels everyone ignores: email and SMS. A phone call converts dramatically better — but no merchant can staff calls at 11pm for a ₹1,400 cart. The economics only work if the caller is software.

02Approach

The agent is built like a good sales rep, not a robot with a script: it knows the cart, the customer's history and the merchant's discount ceiling; it opens honestly as an AI assistant; it handles objections within guardrails; and it knows when to stop. Merchants configure intent, tone and limits — the voice workflow engine handles the rest, from dial to Shopify write-back.

03Architecture notes

  • Trigger. Shopify webhook → eligibility rules (cart value, region, quiet hours, do-not-call) → call queue.
  • Conversation. The MeetAira-derived streaming voice stack, tuned for telephony's 8kHz reality and interruption-heavy dialogues.
  • Guardrails. Discount authority, claim restrictions and mandatory disclosures are enforced in the tool layer, not the prompt.
  • Write-back. Outcomes land in the store: recovered checkout links, applied discounts, tagged customers, transcripts.

04Impact

Zero-touch
From abandoned cart to recovery call
100%
Of calls within merchant guardrails, logged
1 → N
Voice stack reused from MeetAira

05Lessons

Vertical AI products are trust products. The prompt engineering was the easy month; the guardrail engineering — provable limits on what the agent may offer and claim — is the product. Building the second product on the first one's infrastructure is the compounding I optimize for.

06 · Agentic Component Library

One system, two frameworks.

A design system delivered simultaneously to React and Angular codebases from a single source of truth — tokens, behaviour contracts, accessibility and performance budgets included.

Role
Design system lead
Timeline
2023 — 2025
Stack
React · Angular · TypeScript · Design tokens · CI
Status
Adopted across product teams

01Problem

An enterprise mid-migration lives in two worlds: new surfaces in React, a decade of Angular still serving customers. Two implementations of every button meant double the bugs, drifting brands, and accessibility that depended on which year a screen was built.

02Approach

  • Tokens as the treaty. Color, type, spacing, motion and state live in one token pipeline; both frameworks consume generated outputs. Design changes land everywhere or nowhere.
  • Behaviour contracts. Each component has a framework-agnostic spec — states, keyboard map, ARIA contract — and both implementations are tested against the same spec suite.
  • Budgets in CI. Bundle size and interaction latency are budgeted per component; a regression fails the build, not the user.
  • Migration tooling. Codemods map legacy component usage onto the system, making adoption a command rather than a quarter.

03Impact

1
Source of truth for both frameworks
AA
Accessibility contract, tested in CI
0
Visual drift between stacks

04Lessons

A design system is a political artifact wearing an engineering costume. The token pipeline and spec suite mattered because they ended arguments — the system became the place where decisions, once made, stayed made.

07 · Angular Migration Framework

Rebuilding the plane mid-flight.

A framework for migrating a large enterprise Angular application incrementally — with an extraction graph, automated component mapping, and a parity system that proves each migrated screen behaves identically.

Role
Architect & tooling engineer
Timeline
2024
Stack
AST analysis · Codemods · CI verification · Visual diffing
Status
Driving migration in production

01Problem

"Big-bang rewrite" and "never migrate" are both ways of failing — one loudly, one slowly. The codebase was too large to freeze and too entangled to move screen by screen without knowing what each screen actually depended on.

02Approach

  • Extraction graph. Static analysis over the legacy codebase builds a dependency graph — components, services, state, templates — so migration order is computed from coupling, not guessed from opinions.
  • Component mapping. Each legacy component maps to its design-system target with a confidence score; high-confidence mappings are codemodded, low-confidence ones are queued for humans.
  • Verification pipeline. Every migrated screen runs behavioural tests plus visual diffs against its legacy twin in CI.
  • Parity ledger. A live dashboard of what's migrated, what's proven equivalent, and what's intentionally different — the single artifact that kept stakeholders calm.

03Impact

0
Release freezes during migration
Graph-ordered
Migration sequence, not guesswork
Proven
Behavioural parity per screen, in CI

04Lessons

Migrations are evidence problems. The moment parity became something CI proved rather than something engineers promised, the migration stopped being scary — and therefore stopped being slow.

08 · Live Orders Dashboard

Race-day energy for order data.

A realtime order leaderboard inspired by MotoGP timing screens — live rankings, position-change choreography, and enterprise-grade legibility under thousands of updates a minute.

Role
Design & engineering
Timeline
2023
Stack
React · WebSockets · FLIP animation · Virtualization
Status
Live on office screens

01Problem

Realtime dashboards usually fail in one of two ways: they update so fast they're unreadable, or they're throttled so hard they're not really live. This one needed to run on a wall, all day, and make a sales floor glance up when something changed.

02Approach

  • MotoGP grammar. Position numbers, gap-to-leader, and overtake flashes — a visual language people already read at 300km/h, borrowed for orders per minute.
  • FLIP choreography. Rank changes animate as physical overtakes: measure, invert, play. The eye tracks who moved without reading a single number.
  • Update batching. WebSocket bursts coalesce into one choreographed reflow per second — live enough to feel instant, calm enough to stay legible.
  • GPU-only motion. Transforms and opacity exclusively; the board holds 60fps on a five-year-old TV stick.

03Impact

60fps
Under thousands of updates/min
All-day
Wall display, zero babysitting
Instant
Read-at-a-glance rank changes

04Lessons

Delight in enterprise software isn't decoration — it's information design with a pulse. People trusted the board because the motion meant something, every time.