Work
Products built, systems shipped, problems solved
Independent SaaS products and AI systems built from scratch, alongside enterprise-scale analytics platforms delivered at Amazon — a full picture of how I think about data, product, and systems.
AI-Native Aerial Business Platform
Vyoman
An end-to-end, AI-native aerial-photography business — portfolio, print store, ops console, and an LLM + vision model creative pipeline — designed, built, and operated solo. Full order-to-fulfilment pipeline operational.
The system
- Public site + print store (Next.js · Stripe · Gelato POD · DHL tracking · digital licensing · returning customer checkout)
- Canonical data layer in Neon — a dimensional item/variant registry (VSIN) as the single source of truth
- Django control-plane console on Railway — order ops, Gelato lifecycle emails, LLM intelligence dashboard, VLM product matching
- Vantage (vantage.vyomanaerials.com) — 12-phase LLM mission intelligence built in 3 days: streaming LLM-generated briefs (SSE), airspace WMS compliance, weather scoring, shot references via FLUX image generation + Pexels, portfolio ROI linking flights to orders
- ISR propagation: flip an item live in the console and the site updates in seconds, no deploy
Why it matters
The clearest proof of the full arc — canonical data, control plane, intelligence, commerce, and ops — designed, shipped, and operated by one person. Alongside it: an LLM + vision model creative pipeline (footage → graded photos → story draft) and Skyledger, a compliance-first drone-airspace SaaS where a deterministic rules engine decides and AI only drafts.
Portfolio — Luxembourg, from above
Shop — The Three Veins print, €8.00
Restaurant Intelligence Platform · B2B SaaS
Disha Analytics
Compresses 2 analyst-days of restaurant research into a 5-minute structured intelligence report — deployed in production, with pilot reports run on live restaurants.
Problem
Independent restaurant owners across Europe make operational decisions without structured data. They know something is wrong — reviews mixed, margins thin — but cannot afford consultants and have no tool that synthesises insight into action.
What it does
An analyst enters a restaurant name and city. Disha scrapes 8 data sources, runs 6 specialist LLM agents, and produces a ranked action plan, competitor benchmark, hypothesis checklist, and financial signal section — formatted as an internal report or owner-facing walkthrough.
Architecture
- Django 5 REST API + React 18 + Vite frontend
- Celery + Redis async pipeline for report generation
- 6 specialist LLM agents — competitor, financial, review, menu, validation, synthesis
- Three delivery modes: Public Scan, Validated Report, Owner Walkthrough
- Three-tier ingestion layer: format adapters → validators → canonical models
Why it matters
This is a direct translation of the instinct I built over a decade in analytics — taking a complex data problem and compressing it into a decision-ready output. Disha does for restaurant analytics what I did for QBRs and Big Bet BRDs: turns raw data into a clear action plan under time pressure.
Intelligence report — action priorities + brief
Belief graph — structured evidence chain
AML/KYC Compliance Intelligence · Luxembourg · Demo Live
ComplyLens
Compliance intelligence for Luxembourg AML/KYC teams — generates CSSF-inspection-ready review documentation with structured evidence citations, auditable LLM reasoning, and four-eyes workflow.
Problem
Luxembourg AML/KYC compliance teams produce manual review documentation that is difficult to make "inspection-ready" for CSSF audits. In 2024, CSSF handed down €4.7M in fines (up 52%). The problem isn't whether a review happened — it's whether the reasoning is auditable.
Three pillars
- RiskRationale — risk classification with CSSF-mapped evidence citations
- ReviewMemo — 11-section periodic review memo + CSSF audit ZIP export
- AlertDesk — alert disposition with four-eyes approval workflow
Why it matters
Claude LLM drafts structured reasoning with CSSF-mapped evidence citations; humans review and sign. Every determination is auditable by a regulator — not a black-box score. AMLR applies July 2027, creating a first-mover window for tools built on this principle.
Analyst Dashboard — review queue
Client review — Meridian Holdings
Status
Full MVP deployed in 2 days. Password-gated live demo (Meridian Holdings synthetic scenario) at complylens.vercel.app. FastAPI backend on Railway, Next.js 15 frontend on Vercel. Pre-seed exploration with co-founder Sachin Bhalerao; targeting Luxembourg ManCos, AIFMs, and compliance boutiques.
Community Trust Infrastructure · Deployed
CircleWorks
Trust infrastructure for the informal local economy in residential societies — workers build verifiable trust scores, residents post jobs and receive LLM-powered pricing, money flows directly peer-to-peer.
Problem
The informal labor market inside residential societies (plumbers, electricians, cleaners) operates entirely on word-of-mouth with no trust signals, no transparent pricing, and no accountability. Residents cannot find reliable workers; good workers cannot build a reputation that travels.
What it does
A resident posts a job via the web portal. An NLP intent parser extracts job details — skill, duration, headcount — and an LLM pricing engine suggests a rate range. Matched workers receive alerts on Telegram, accept with one tap, complete the work, and build a verifiable trust score. All payments are peer-to-peer via UPI — CircleWorks never touches money.
Architecture
- Django 5.1 backend with PostgreSQL on Railway
- Telegram bot for worker onboarding, job alerts, and confirmations
- Unified web portal — resident, worker, and admin views
- NLP intent parser + LLM pricing engine (Claude API) on every job
- Trust score built from completions and resident reviews
- Platform never handles money — all payments direct via UPI
Type what you need — AI extracts the rest
LLM pricing engine — ₹200 recommended
Status
Platform built and deployed to Railway. Pilot setup complete at Blueberry Homes, Bangalore — worker onboarding and ops documentation ready. Full SOPs, dispute resolution, and field ops authored in collaboration with on-ground partner Alok.
Privacy-First · Local LLM · Live at soulspark.me
Soul Spark
A multi-agent LLM life intelligence system that integrates finance, health, career, and growth signals into a unified conversational advisor — all inference local via Ollama, no personal data leaves the device.
Problem
Most personal data systems are fragmented — health, money, tasks, and behavior live in separate apps, making it difficult to turn data into clear decisions. Cloud-hosted LLM inference services require routing sensitive personal signals through an external provider.
System Design
- Local Ollama LLM inference — all data stays on-device
- Packet-native multi-agent architecture with internal graph runtime
- Shared memory and pending-action loop across agents
- Finance, Health, Career, Social, Growth, Library domains unified
- Cloudflare tunnel for public access via soulspark.me
Why it matters
It reflects my interest in moving beyond analytics dashboards toward systems that actually help users make decisions. The shift from "here is your data" to "here is what you should do" is where LLM-powered reasoning creates real leverage — and where privacy-first design matters most.
AI-Native Audiobook System · Daily Driver
Narrate
Converts any document into a narrated audiobook — voice personality per book, AI cover art, adaptive streaming, and lock-screen captions. Built for daily personal use. Designed for anyone who prefers listening — and for people facing barriers to reading through vision, literacy, language, or attention.
Problem
Reading backlogs never shrink — PDFs and whitepapers sit unopened because sustained reading requires uninterrupted attention. Audiobook platforms only serve commercial content. The same document-first sources are completely inaccessible to anyone who cannot read due to visual impairment, low literacy, or language.
Architecture
- Narrate-Ready pipeline — 9-rule preprocessing spec for any document type
- Three-layer narration quality: deterministic preprocessing + voice personality + structural Narrate-Ready
- Adaptive buffered conversion — 200-chapter book starts in under 30 seconds
- AI cover art (DALL-E 3) · intro/outro bookends · "Where you are" context · narrated brief
- 3-mode lock-screen captions via MediaSession API · PWA
- No database — Cloudflare R2 as complete data layer
Why it matters
Narrate proves product craft in a different way: not only architecturally complex, but useful enough to become a daily habit. Eight books narrated, including a 125-chapter academic text on AI strategy. The Narrate-Ready spec is a publishable preprocessing standard — and the same architecture extends naturally into universal audio accessibility for public notices, official documents, and multilingual narration.
Status
Live at narrate-psi.vercel.app. Used daily — 8 books narrated to date. Next.js 15 + OpenAI TTS + DALL-E 3 + Cloudflare R2, deployed on Vercel. Single-user, privacy-first, no subscriptions.
Vyoman · AI Geospatial
Vyoman Earth Intelligence
Turns Luxembourg's public geospatial data and drone imagery into an evidence-backed, human-reviewed site briefing — in minutes, not a GIS analyst's week.
Problem
Site analysis in Luxembourg means manually querying three separate government WMS servers, interpreting LiDAR elevation data, and cross-referencing drone imagery — a GIS analyst's week per site. No tool assembles this automatically for the Luxembourg data stack.
Architecture
- Live Geoportail WMS orthophotos (10cm source resolution) + LiDAR DEM + Vyoman drone imagery
- Evidence-packet architecture — the same pattern as ComplyLens, Disha, and Soul Spark
- 2D map + 3D PyDeck terrain renderer side by side
- Human review workflow: draft → reviewed → approved + DuckDB audit trail
- Local-first Ollama inference + deterministic mock fallback, zero cloud API keys
- ReportLab PDF export with embedded maps, citations, and reviewer annotations
Why it matters
Proves the evidence-packet architecture generalises beyond documents and restaurant reviews — to live geospatial data. Every Vyoman drone flight feeds directly into this pipeline, closing the loop between the aerial operation and the intelligence tool. The natural next step: Vantage mission data and site briefings from the same canonical data layer.
Status
MVP 0 — local prototype. Streamlit + Python 3.13 + DuckDB + GeoPandas + Rasterio + PyDeck + Ollama + ReportLab + Geoportail WMS. Extends the Vyoman Aerials drone operation. No public deployment yet.
Three independently deployed LLM-powered tools built on Python + Streamlit — each solving a concrete user problem with LLM inference layered on structured data. All live on Streamlit Cloud.
Vivre Ensemble
Luxembourg citizenship exam study platform and mock test engine
Problem
The Luxembourg "Vivre Ensemble" civic knowledge exam covers dense constitutional, historical, and institutional content — with no structured study tool or practice exam available.
What it does
A zero-dependency, offline-capable single-page web app with three modes: a structured study guide (40+ exam topics), a timed 40-question mock exam drawn from 80+ questions with instant explanations, and a civic history narrative for long-term Luxembourg residents.
Why it matters
Built for personal exam prep. The fastest path to a high-quality study tool was building it from scratch — a useful exercise in pure web performance without framework overhead.
Skyledger
In Development · EU PilotCompliance-first EU drone airspace intelligence — deterministic rules engine, PostGIS zones, immutable audit trail
Problem
EU drone operators navigating EASA airspace must comply with a complex overlay of geographic restrictions, category-specific authorisations, and frequently updated national rules. Existing flight-planning tools are not compliance tools — they do not audit a flight plan against current regulations or produce defensible records.
What it does
A PostGIS-backed airspace intelligence platform: EU zone geometry served as MapLibre vector tiles for real-time zone visualisation, a deterministic versioned rules engine that evaluates flight plans against EASA Open/Specific/Certified category rules, and immutable audit records for every determination. AI drafts supplementary guidance; all legal compliance outcomes are decided by the rules engine alone.
Why it matters
The core design constraint — AI drafts, rules engine decides — matters because airspace determinations carry legal weight. A deterministic, version-controlled rules engine makes every compliance decision traceable and auditable, not a black-box output.
Analytics systems, data products, and infrastructure delivered during 9 years at Amazon — spanning payments, private brands, and global marketplace expansion across India, Luxembourg, and multiple regions.
Selection Expansion Analytics Platform
Amazon · 2024 – 2026Amazon · Luxembourg · L6 · 17+ global stores
Problem
17+ global Amazon stores had fragmented, untrustworthy selection data. There was no reliable single view of what was expanding, what the funnel looked like, or how to frame it for leadership planning cycles.
What I built
- Served as BI functional lead, advising Director-level monthly business reviews with the insights behind key strategy decisions for a ~$4B GMS cross-border seller program across 17+ global stores.
- Redesigned the metric taxonomy — seller sales, offer quality, Contribution to 3P, selection quality — and rebuilt the QBR framework adopted across 4 senior leadership planning cycles.
- Restored data integrity across an $838M+ sales program by rebuilding core seller identification logic, restoring trust in 20+ downstream reports used daily by Business Managers and engineering teams.
- Authored BI Requirements Documents for two Big Bet products ($11.1B addressable GMS): Listing Preferences, the platform migrating Amazon's cross-border listings (32-day sprint, Dec 2025 global rollout), and Lighthouse, a new seller segment onboarding ecommerce platform integrators previously unsupported at Amazon, approved by Amazon Europe VP.
Private Brands BI Platform
Amazon · 2021 – 2023Amazon · Luxembourg · L5→L6 · EU/NA/APAC
Problem
EU Private Brands had no analytics function. Product, Marketing, Supply Chain, and Finance were all making decisions without reliable shared data, tooling, or a common metrics language across regions.
What I built
- First BIE hire for EU Private Brands — built the BI function from scratch and formally managed 5 direct reports (3 BI Engineers + 2 interns), owning the full analytics stack across four business domains.
- Delivered the APB-first Deals Performance Dashboard: 6 views, 20+ metrics, 90+ users across 10 countries and 12 job families. Global single source of truth, saving 350+ hours/year.
- Built Catalog Defects and products-with-a-sale bridge analysis uncovering ~€110M GMS opportunity — adopted by senior leadership for quarterly planning.
- Designed and led Inventory Tracker Engine with global supply chain teams, saving 1,000+ hours/year; extended to NA and APAC in V2.
- SQL cube innovation reduced the weekly business review codebase from 20K to 1.3K lines — adopted org-wide across APB BI.
Amazon Pay & HFC Analytics System
Amazon · 2017 – 2021Amazon · Bengaluru, India · L4→L5 · Payments & Commerce
Problem
Amazon Pay and High-Frequency Commerce (Recharges, Bill Payments, Flights) needed analytics infrastructure built from scratch — for product launches, marketing campaigns, fraud prevention, and Go/No-Go experiment decisions, all simultaneously.
What I built
- Built end-to-end data pipelines, 5 dashboards with 15+ views, and a 25+ query self-serve library supporting the launch of 8 HFC categories — foundational analytics for a $2B business driving 50+ launch campaigns.
- Designed the Downstream Impact methodology for product launch decisions — Go/No-Go framework for 5 Pay experiments at $30M GMV stakes, adopted as the PM standard for feature launches.
- Built a Python-based self-serve customer segmentation tool processing ~100 marketing campaigns/month, reducing campaign setup time by ~70% and saving ~3,000 hours/year.
- Designed abuse-prevention rules combining multiple data sources to auto-cancel ~100K abusive HFC orders/month (~2% of volume), significantly reducing incentive fraud with minimal false positives.