staff software engineer · biva · pre-series A00:00:00 · local

SYSTEM · 11 YEARS IN PRODUCTION

An operating system for turning ideas into software.

I'm Akhil Kapuagnti — Staff Software Engineer at Biva. Eleven years from hand-rolled infinite scroll over 150K records, through airline-scale frontends, to founding-engineer platform builds and production AI infrastructure. This isn't a portfolio; it's the live system behind the work.

Enter the system Resume

“Software is just memory with intent — structure, recall, and the will to act on it.”

— working note · on what software is

MEMORY

The concepts the system is built on.

// hover a node to trace its relationships across the graph

AI Agents

Orchestration

MCP

Retail Ontology

Architecture

Voice AI

Frontend Performance

PROFESSIONAL SYSTEMS

Not case studies. Deployed systems.

Built at Biva, a pre-Series A retail analytics company where I own the application end-to-end — read the way they were built: problem, architecture, tradeoffs, returns.

node.biva-neural

BIVA Neural

Conversational AI layer over retail data

PRODUCTION · ACTIVE

PROBLEM

Retail teams sat between their questions and their data — every insight needed an analyst, a dashboard, or a ticket. Reports over billions of rows took too long to be part of a conversation.

TRADEOFFS

Built the AI layer on top of our data team's Trino query engine (their work took a flagship report from 27 min to under 2) rather than owning retrieval myself — clean seam, each layer debuggable alone. Structured intent over free-form SQL generation: less magic, more trust.

ARCHITECTURE

NL QuestionAI OrchestratorTrino · Retail OntologyInsight · Auto-PPT

RESULTS

900

enterprise clients reached

<2 min

reports that took 27

Agent

next on the roadmap

node.mcp-infra

MCP Infrastructure

Client-facing data access protocol

SHIPPED · IN USE

PROBLEM

An enterprise client wanted their own AI tools to talk to BIVA data directly — not through our UI. Ad-hoc APIs would mean reinventing auth, context and error handling per integration.

TRADEOFFS

Protocol overhead up front for composability later. I built the full MCP stack — authentication, orchestration and developer tooling — once, as infrastructure, instead of bespoke glue per client.

ARCHITECTURE

Client AgentMCP ServerAuth · OrchestrationBIVA Data

RESULTS

Production

client-driven, in use

Typed

tool contracts

Once

auth + errors solved

node.biva-platform

BIVA Platform

Retail analytics · Halo → BIVA

DEPLOYED · SCALING

PROBLEM

Our D2C analytics product (Halo) had to pivot to enterprise retail. Patching wouldn't survive the ontology change — retail data means a different shape of everything.

TRADEOFFS

Re-architected from the ground up instead of patching — expensive quarter, but the app has survived 2–3 major version upgrades since and still runs smooth. Owned everything app-side (auth via Keycloak, middleware, frontend) while the data science team owned models.

ARCHITECTURE

AngularNestJS MiddlewareKeycloakRetail Ontology

RESULTS

10 live

retailers, ₹3,000cr scale

0→1→10

now scaling 10→100

2–3 eng

team I oversee

SIDE SYSTEMS

Built after hours. Deployed anyway.

Sole architect on all four — each one exercises a different discipline of the system.

node.donna

DONNA

BETA

AI SYSTEMS & AGENT ORCHESTRATION

Voice-first personal assistant with a privacy-first, local-first architecture. LangGraph + GPT-4o tool calling parses voice into structured intent across health, tasks, finance and system control; a macOS agent executes on-device, the cloud relay stays a thin pass-through.

LangGraphGPT-4oDeepgramFastAPIReact Native

in beta testing · multi-agent pipeline · intent JSON

node.ipolens

IPO Lens

WORKING

MULTI-MODAL PIPELINES & CONSTRAINED GENERATION

End-to-end pipeline turning 400+ page IPO prospectuses into publication-ready 60-second YouTube Shorts. Gemini multi-modal extraction of financials and risk factors; constrained script generation with word-count enforcement and compression passes; Remotion renders data-driven 9:16 video.

Gemini 2.0OpenAIRemotionNext.jsSupabase

400+ page PDFs in · timed 60s narration out

node.church

Church Platform

LIVE · 2 CHURCHES

ENTERPRISE SAAS & SYSTEM ARCHITECTURE

Consolidates members, finances, attendance and communication — replacing spreadsheets and disconnected tools — with automated notifications, reporting and role-based access. Built and operated as sole developer.

sole developer · in productionholyzioncsichurch.in ↗

node.incent

Incent

LIVE

PRODUCT ENGINEERING & USER EXPERIENCE

Digital gifting platform for buying and distributing gift cards — business onboarding, secure transactions, and an optimized gifting flow for both businesses and consumers. Built from scratch with one collaborator.

co-built · recently launchedincent.store ↗

ARCHITECTURE

How the systems flow.

Three systems I architected end-to-end — one professional, two personal. Lines show what talks to what.

// system.biva-neuraltrino engine · retail ontology · auto-PPT

Input

NL QuestionChat UI

AI Orchestrator

Intent ParserQuery Planner

Data Layer

Trino EngineRetail Ontology

Output

InsightAuto-PPT

natural-language question → orchestrator plans → trino queries retail ontology → grounded insight, chart or PPT out

OPERATING PRINCIPLES

The defaults the system runs on.

Hard-won over 11 years and four companies — the bias every project inherits before its specifics.

Separate memory, reasoning and action.

If they share one prompt, none of them can be trusted or scaled.

Choose boring tech you can defend.

Picked Angular + NestJS when the world was obsessed with React. Five years and one pivot later, it's still shipping.

Re-architect when the ontology changes.

You can patch features. You cannot patch a wrong model of the world — D2C to retail taught me that.

Standardise the boring parts.

Auth, tooling and observability are infrastructure, not features — solve them once. That's why MCP, why Keycloak.

Design for the client you don't have yet.

One enterprise contract can mean 900 downstream clients. Isolation and scale are cheaper built in than bolted on.

Ship the smallest honest system.

Then let real use show you where it actually needs to grow.

RESEARCH

Active investigations, not blog posts.

Open questions the system is still working through — each with a hypothesis and a state.

ACTIVE

RT-01 · this week

From talk-to-data to a true analytics agent

BIVA Neural answers questions and builds PPTs. Can it plan, verify and act on retail data without a human in every loop?

current lean · structured intent first; autonomy earned one verified step at a time

55% explored

ONGOING

RT-02 · weekly

Scaling 10 → 100 retailers without re-architecting

Onboarding steps up every week. Which seams hold — and which assumptions from the 0→10 era quietly break?

current lean · the ontology holds; operational tooling is the real bottleneck

35% explored

ACTIVE

RT-03 · beta

What voice changes about assistant UX

DONNA's beta testers don't type — they talk. How much structure can you recover from natural speech before users feel interrogated?

current lean · intent JSON from one utterance works; multi-turn repair is the hard part

60% explored

VALIDATED

RT-04 · shipped

MCP as a production contract, not a demo

Does a typed tool protocol survive a real enterprise client's requirements — auth, orchestration, developer tooling included?

current lean · concluded — shipped at Biva, in production use by a client

100% · concluded

DECISION LOG

What I decided, and why.

Real architecture decisions from 11 years — the call, the reasoning, and how it aged.

DL-021

context · biva · 2025

MCP as the contract for client data access

One typed protocol with auth, orchestration and tooling solved once — instead of a bespoke API per enterprise client.

ADOPTED

DL-018

context · biva neural · 2025

Build the AI layer on the data team's Trino engine, not around it

Their 27min→2min query work stays theirs; my layer orchestrates on top. Clean ownership seam, honest credit, independent scaling.

ADOPTED

DL-015

context · donna · 2024

Local-first execution; the cloud is a thin relay

Device-side agent handles execution so personal data never has to leave. Harder distribution, but privacy by architecture — not by policy.

ADOPTED

DL-009

context · biva · 2022

Re-architect Halo for retail instead of patching the D2C model

The ontology changed, so the architecture had to. Cost a quarter; the app has survived 2–3 major upgrades since without a rewrite.

ADOPTED

DL-004

context · saras · 2021

Angular + NestJS over React, against the fashion

Huge enterprise forms, lazy loading, dependency injection — the boring stack fit the problem. Still shipping five years and one pivot later.

ADOPTED

DL-001

context · tripod · 2017

Hand-rolled infinite scroll for 150K records, pre-framework era

No virtual-scroll libraries existed for us — so I built windowed rendering in plain JavaScript. The screen went from unusable to instant.

ORIGIN

EVOLUTION · 11 YEARS

How the system learned to build.

2015 — 2018 · TRIPOD

Foundations

Junior on eureqaa — requirements, product thinking, and performance craft: windowed rendering for 150K records in plain JS. The product was later acquired.

2018 — 2020 · ATMECS × JETBLUE

Scale

Rebuilding JetBlue software for Mission 2020 — Angular in depth, first paint, lazy loading, tree shaking. First time leading a team of 4–5.

2020 — 2022 · SARAS

Ownership

Founding member of Halo — owned frontend and middleware from repo zero. First full architecture calls, integrations, and the finance side of building.

2022 — NOW · BIVA

Staff · AI Systems

Own the application end-to-end through a pivot and re-architecture. Shipped MCP infrastructure and BIVA Neural. 0→1→10 done; scaling 10→100.

SYSTEM ONLINE · ACCEPTING CONNECTIONS

Open a channel to the system.

For staff/senior roles and AI product work. Tell me the problem; we'll design the architecture.

[email protected]