I build the user-facing layer of AI products —
turning model capabilities into usable,
reliable, and polished interfaces.
Featured Case Study
Sahabat-AI by Indosat
Indonesia's first national-scale AI platform — a 70-billion parameter open-source LLM
serving 2M+ users across 5 languages, backed by one of the country's largest telecoms.
Context
Sahabat-AI is a sovereign AI initiative by Indosat Ooredoo Hutchison (~100M subscribers) in partnership
with GoTo Group, AI Singapore, Google, and Nvidia. The platform supports Bahasa Indonesia plus 4 regional
languages (Javanese, Sundanese, Balinese, Bataknese) — keeping the model, data, and infrastructure within
national boundaries.
Two phases, two layers of the stack. Phase 1: architected the Python orchestration layer between
the platform's frontends and the model routing infrastructure. Phase 2: built AIRIM
(AI Ready in Minutes) — the B2B portal that lets any business deploy a chatbot powered
by Sahabat-AI. Deployed on-site in Indonesia for stakeholder discovery with Indosat's product and
engineering teams.
• Architected the LLM orchestration layer in Python —
the middleware handling request routing, response streaming, and model-switching logic between the
platform's frontends and the model routing infrastructure.
• Built AIRIM end-to-end — the self-service B2B
portal inside Indosat's existing Digital Partnerships portal: business registration, website
scraping for RAG, document upload and ingestion, API key management, a PlayField testing sandbox,
agent registration, and the full payment flow.
• Created the embeddable chat widget — the HTML
script businesses drop into their own site to get a working chatbot that answers from their
scraped content, streaming responses token-by-token.
• On-site stakeholder discovery — flew to Indonesia
for requirements gathering with Indosat's product and engineering teams.
Technical Challenges
Multi-model traffic balancing
The platform routes between Sahabat's homegrown 70B model and frontier model fallbacks. The
orchestration layer abstracts model-switching so the frontend renders responses identically
regardless of which model generated them.
Streaming inside someone else's website
The AIRIM widget renders responses token-by-token as the model generates — inside a customer's
own website, where you control nothing about the host page. Built streaming response handling
that stays smooth without blocking the main thread or breaking the surrounding site, even on
lower-end Android devices common in Indonesia.
Low-bandwidth optimization
Indosat's 100M+ subscriber base includes users on 3G mobile connections across the Indonesian
archipelago. Every interface decision — bundle size, API payload compression, progressive loading —
was shaped by network reality, not lab conditions.
Self-service RAG onboarding
AIRIM turns an arbitrary business website into a chatbot knowledge base with no engineer in the
loop: URL validation, page discovery and selective scraping, document upload and ingestion, and
storage entitlements — designed so a non-technical business owner can go from registration to a
working chatbot in minutes.
Architecture
User requests flow through the React frontend to the orchestration layer, which routes to either
Sahabat's homegrown 70B model or a frontier model fallback. Responses stream back token-by-token
for real-time UI rendering.
The Pattern, Live
The token-streaming pattern from the AIRIM chat widget — the one businesses embed in their
own sites — rebuilt here in ~3 kB of vanilla JavaScript. Scripted responses: no API,
no network requests, nothing leaves this page.
Keyboard accessible · Respects reduced motion
streaming-demo · scriptedidle
AI
Hi — this is a scripted demo of the streaming pattern from AIRIM, the B2B chatbot
platform Sanket built on Sahabat-AI. Pick a question below, or type your own.
Outcomes
2M+Platform Users
70BParameters
5Languages
100M+Indosat Subscribers
Covered by Fortune Asia. Endorsed by Indonesian government ministers at the national launch.
Partnership ecosystem includes GoTo Group, AI Singapore, Google, and Nvidia.
A major healthcare organization needed a patient-facing platform that met strict WCAG 2.1 AA
accessibility requirements while serving 100K+ users. I led the frontend implementation with a
component architecture where every interactive element was built accessible-first — ARIA roles,
keyboard navigation, and screen reader support tested at the component level, not the page level.
Mentored 5 engineers on accessibility patterns, improving sprint velocity by 25%.
99.9% uptime
Banking
Payment Gateway Redesign
A Fortune 100 banking client needed a complete redesign of their payment gateway UI — every pixel
has compliance implications and every interaction carries financial liability. I led the frontend
redesign within PCI DSS constraints dictating input masking, session management, and data handling.
The architecture prioritized progressive enhancement — if JavaScript failed to load, the payment
flow still worked.
+15% engagement
Semiconductor
ML Annotation Tool
A major semiconductor company needed internal tooling for labeling and tagging datasets used in ML
training pipelines. I built the annotation interface — a data-dense, high-interaction tool where
annotators spend 8+ hours daily. The design focused on keyboard-first workflows, minimal
context-switching, and batch operations that reduced per-annotation time. Enterprise internal tooling
where the UX challenge is efficiency, not discoverability.
-40% per-annotation cycle
Government
New York State POC
Through Accenture's government consulting practice, I developed a proof-of-concept application for
New York State/City government. Public-sector work operates under constraints commercial projects
don't face — Section 508 accessibility mandates, specific security requirements, and stakeholder
review cycles that add weeks to every decision. Navigated those constraints to deliver a functional
POC that secured stakeholder approval to proceed to full implementation.
Approved for production
Media / Streaming
Video Streaming Platform
A media company needed frontend modules for their video streaming platform — content discovery UIs,
playback controls, and adaptive streaming interfaces. The technical challenge was building UIs that
respond to variable network conditions in real time: adjusting stream quality indicators, preloading
content based on user behavior, and rendering media-heavy layouts without blocking core interactivity.
Optimized for CDN-aware asset loading and service worker caching.
<2s time-to-first-frame
Side Projects & Open Source
Technical experiments and developer tools built outside of client work.
VS Code Extension
SimpleSync
A technical experiment in peer-to-peer architecture — real-time LAN file synchronization for
VS Code using WebSocket and mDNS discovery. Built to explore P2P patterns and offline-first
networking, not as a flagship product.
55+ developer utilities with a privacy-first, client-side-only architecture. No server,
no tracking, no accounts. Every computation runs in the browser — data privacy guaranteed
by architecture, not policy.
CSV/Excel/JSON chart maker where 100% of processing happens client-side. Designed for teams
handling sensitive data who can't risk uploading to third-party services. Privacy guaranteed
by architecture — zero network requests for data.
Every interface I build assumes 3G speeds and intermittent connectivity. Building Sahabat-AI for
Indonesian mobile users on varied networks taught me that performance isn't a feature — it's the
feature. If your AI product takes 8 seconds to render a response, users don't care how smart the
model is.
02
Accessibility is architecture, not a checklist
WCAG compliance can't be bolted on after launch. On the healthcare platform, accessibility was the
load-bearing wall — 100K+ users with varying abilities depended on it. I treat ARIA roles, keyboard
navigation, and screen reader support as architectural decisions, not QA tickets.
03
The best orchestration layer is invisible
On Sahabat-AI, we swapped underlying LLM models mid-sprint. The frontend needed minimal changes —
a configuration update, not a rewrite. That's the test: if your orchestration is well-abstracted,
the team shipping UI shouldn't need to know or care which model is running underneath. Good
infrastructure makes itself invisible.
Leading frontend architecture for enterprise AI and healthcare products at scale.
• Led end-to-end delivery of an AI application for
Indosat (Indonesia), from on-site stakeholder discovery through production launch for a platform
serving 2M+ users.
• Architected LLM orchestration layer in Python and
built high-performance React interfaces for RAG workflows, connecting AI pipelines to intuitive
user experiences.
• Shipped WCAG 2.1 AA-compliant healthcare platform
for 100K+ users with 99.9% uptime. Mentored 5 engineers, improving sprint velocity by 25%.
09/2021 — 04/2024
Accenture
Custom Software Development Analyst
Shipped component libraries and performance-optimized UIs across enterprise client engagements.
• Built comprehensive component library with 50+
reusable React components leveraging hooks, custom hooks, and compound component patterns,
reducing development time by 30%.
• Redesigned payment gateway UI from the ground up,
increasing user engagement by 15%. Maintained 85%+ code coverage across all projects.
• Optimized frontend performance — 35% reduction in
bundle sizes through code splitting, lazy loading, and tree-shaking, improving Core Web Vitals
across all client projects.
09/2019 — 08/2021
Applied Cloud Computing
UI Developer
High-performance web applications using React and Next.js.
• Built and optimized UI components and frontend APIs,
improving application load times by 40% through performance profiling and optimization.
• Implemented state management solutions using Redux
and Context API for complex application workflows.
• Integrated RESTful APIs and handled asynchronous
data flows with modern patterns (React Query, SWR).
Sanket is a phenomenal UI expert who also possesses a deep understanding of backend architecture. His
ability to anticipate how frontend decisions impact backend scalability is rare and invaluable.
"
The modular component systems he architected saved us months of technical debt. He doesn't just build
beautiful interfaces; he builds full-stack systems that are robust and incredibly fast.
"
Sanket seamlessly bridges the frontend-backend divide. His optimizations improved our app's
performance by 40%, and his mentorship elevated the entire team's engineering standards.