Meet Vaghani

Full-stack engineer building LLM agent infrastructure — an MCP server with a custom OAuth 2.1 Authorization Server, the platforms underneath it, ADK and Vertex AI agents, and the complete AI layer inside a live trading platform. Also at data scale: a 35 TB migration with 100% verified integrity.

Currently

SDE-1 at DevX AI Labs

0 TB
Migrated
0K+
Orders reconciled
0+
Features shipped
0
Security findings
0+
Products indexed
  

The calendar is catching up. The shipping list isn't slowing down.

Selected work · 2026

What I've shipped.

Built, measured, in production.

Agent infrastructure, data platforms and live products across DevX AI Labs, GlideTech Solutions and Escape Plan. Scroll to step inside each one.

01 / 16MCP server· DevX AI Labs · Cloud Run + Terraform

Pulse MCP

Claude-native access to a whole delivery platform.

Two halves. A remote Model Context Protocol server giving every employee Claude-Code-native access to a delivery-governance platform through 35 RBAC-scoped tools across 13 domains, with a custom OAuth 2.1 Authorization Server written because MCP clients require Dynamic Client Registration and no managed provider offered it. And the product-side facade it talks to, which checks per-tool permissions and writes an audit entry on every call — because the broker is deliberately not allowed to decide anything itself.

  • 35 RBAC-scoped tools, 13 domains
  • Custom OAuth 2.1 AS · DCR · PKCE · JWKS
  • Facade re-derives access on every call
  • Stateless broker, zero authority, full audit
dossiers 15 + 16
02 / 16AI subsystem· GlideTech · Flask/FastAPI · EKS + ArgoCD

TradeScanner AI

The entire AI layer of a live trading platform.

Signal explainer, trade-plan builder, streaming conversational coach with tool-calling, document and vision analysis, a pluggable RAG stack and a trading journal — 71 modules behind a two-tier cost-routed model layer, swappable to OpenAI with one environment switch.

  • LLM compliance classifier on every output
  • Refuses instead of hallucinating — noData + asOf
  • 26s → 1.5s cold, 1.35s → 9ms warm
  • Paper and live books separated at the model
dossier 40
03 / 16Web app · sole author· GlideTech · Next.js 16 · React 19

TradingLogic AI

Empty repo to shipped product in one week.

A trader-coaching product built from nothing in a single week: signal explainers, trade plans, a streaming coach and a journal, on a design system built by hand rather than pulled from a UI library — with a full GSAP motion system and React Three Fiber scenes including a 3D equity surface, each carrying a designed static fallback.

  • Whole product, start to ship, in one week
  • 22-primitive design system, no UI library
  • GSAP ScrollTrigger · Flip · SplitText
  • Provenance and failure states built in
115 features · 12 areas
04 / 16Cloud migration· DevX AI Labs · Alibaba OSS ⇄ GCS

Music Tribe

35 terabytes moved without losing a byte.

A distributed migration pipeline moving 35 TB across 5.6M+ files with parallel workers and streaming transfers — zero downtime, a 100% verification match, zero errors. Fingerprinting found 944,346 duplicate groups, but nothing was deleted until every one had been checked against the live CDN: 70,038 byte-identical files were still being served at their exact paths, and deleting them would have been an outage. Plus a SharePoint → Azure Blob track for a 1,700-site estate, where measurement beat intuition: 50 concurrent workers failed 26 of 58 transfers where 20 failed 5 of 38.

  • Every duplicate checked against the live CDN first
  • ~60% dedup, 8.85 TB reduced, zero egress
  • Two independent verifications, concurrent
  • Proved more parallelism was worse
dossier 10
05 / 16Data platform· DevX AI Labs · Pimcore on GCP · Terraform

Concordance

Where scattered product data stops disagreeing.

Digital assets, product information and master data reconciled onto one platform for a global audio-hardware group — Pimcore on Compute Engine and MySQL behind a VPC, with IAM service accounts, load balancers and GKE. Brand data arrived as unstructured spreadsheets, every source disagreeing with the next about the same product, and left as normalised relational models resolving to one authoritative record apiece. The reporting layer was partitioned and clustered across millions of rows in BigQuery and deliberately shaped so that it could later ground a retrieval pipeline — which is precisely what was built on it afterwards.

  • DAM, PIM and MDM on a single platform
  • Spreadsheet brand data into relational models
  • Shaped as RAG grounding before anything asked for it
  • Terraform end to end, cloud-agnostic on purpose
dossier 10
06 / 16Web platform· Escape Plan · FastAPI · AWS · live

CarryScore

A luggage comparison platform that scores honestly.

Tech lead on a live platform indexing ~4,500 products across 9 brands and 20+ specs. The catalogue is built from Postgres at boot and held in RAM, with live prices painted onto it and a pre-gzipped fast path for the unfiltered request — and if the database is unreachable it refuses to start rather than falling back to the bundled file.

  • 0–100 algorithm, 5 dimensions, 6 profiles
  • 15,291 reviews analysed with Claude
  • Scores suppressed below 10 reviews
  • 4,518-page SEO audit, then the rebuild
mycarryscore.com
07 / 16ML engine· DevX AI Labs · FastAPI + Python

Foxtale Forecast

14% error where the naive baseline gave 27%.

An end-to-end SKU-level demand-forecasting engine — ingest, stockout and sale-day cleaning, weekly forecast combination, channel allocation — powering a ₹3.2 Cr / 63k-unit monthly plan across 11 channels. Now self-learning: adaptive conformal inference widens the interval from observed coverage, and a champion/challenger loop promotes a model only on a leakage-free backtest win.

  • 14% monthly WAPE vs 27% naive
  • Adaptive conformal + champion/challenger
  • Projections labelled scenario, not data
  • Reported honestly when ML lost to naive
dossier 12
08 / 16MCP gateway· DevX AI Labs · AWS Lambda + Amplify

Hiring MCP

Apply for a job through your own AI agent.

A stateless MCP hiring gateway letting candidates apply through their own AI agent — zero data at rest, one-time-token lifecycle, write-through and one-application-per-job integrity, and PDF validation by magic bytes.

  • Zero data at rest
  • Split by trust boundary, not convenience
  • Server extracts contact-critical fields
  • LLM resume-vs-JD screening
dossier 17
09 / 16Hiring platform· DevX AI Labs · Next.js 16 · Bedrock + Vertex

Atlas

Every hire, application to probation, in one system.

A hiring and onboarding platform that carries a candidate the whole way — application, screening, scheduling, rounds, offer, onboarding, probation — with the scheduling done by a link the candidate books themselves and the rounds pulled straight off the calendar. Three AI paths run inside it — interview feedback drafted from the Meet transcript, résumé extraction and de-duplication on upload, and a fit score that drives the candidate filters — with the recruiter-facing assistant a system of its own alongside them.

  • SDK retries off: one send is one billed call
  • Traces carry HMAC pseudonyms, never raw ids
  • AI drafts the feedback; the reviewer still edits
  • Calendar watch channels · Meet artifacts on Pub/Sub
dossier 14
10 / 16In-product assistant· DevX AI Labs · ADK on Vertex AI Agent Engine

Hiring Assistant

Ask for the person, not the filter.

Recruiters search the talent pool by saying what they want — senior backend engineers, who knows RAG, anyone AWS certified — instead of stacking filters. The split underneath is deliberate: retrieval recalls a shortlist and applies only the objective facets, seniority and scope, then the model decides who actually fits, because it is the part that knows Java is not JavaScript and that a LangChain profile answers a question about RAG. A coordinator routes between finding people and going deep on one, and a follow-up resumes the specialist you were already talking to rather than going back through the front door.

  • Code filters the facts, the model judges meaning
  • Skill badges are display, never a relevance gate
  • Returning ids only cut a 2,400-token call per turn
  • If retrieval fails it says so, it never invents
dossier 18
11 / 16Delivery platform· DevX AI Labs · Next.js 16 · Postgres + Bedrock

Pulse

Project health, and a matcher that declines to guess.

A delivery-governance platform: a RAID register, action register, project plan, task board, weekly reviews scored against a weighted health matrix, CSAT campaigns and a portal the client logs into. Underneath it sits a matcher that files every calendar meeting against the right project from seven signals — embeddings, vector search, an LLM reranker, learned matches, name, content and Drive links.

  • Cosine and rerank never share a threshold
  • Below the floor it returns nothing, not a guess
  • Every manual assignment teaches the matcher
  • Health matrix scores the weekly review
dossier 15
12 / 16ADK agent· DevX AI Labs · Vertex AI Agent Engine

Emissary

Authority it is lent, never authority it holds.

A Google ADK agent on Vertex AI doing reads and writes across RAID, actions, tasks, plans and weekly reviews — built end to end on the premise that a language model must never be trusted with identity. It states neither who it is nor whose records it may touch. Authority arrives instead as an HMAC-sealed bearer credential carrying an employee id and nothing else — no project list, no role — and every item is addressed by an unforgeable capability rather than an id: an opaque ref that designates the record and confers the right to it in a single token, and fails closed on a wrong kind or a tampered signature. The product re-derives project scope and per-feature role on every call, so nothing the model asserts about itself can widen what it reaches.

  • Opaque HMAC refs — raw ids never reach the model
  • PII scrub before every tool result is seen
  • Names resolved server-side, 400 on ambiguity
  • Router collapsed: the handoff cost 2.8s for nothing
dossier 18
13 / 16Assessment platform· GlideTech · FastAPI · Next.js · Lambda · MCP

ProofHire

The machine assembles evidence. A person decides.

A multi-tenant hiring-assessment platform across three services — an API, a web client and an MCP server — where the candidate gets the coding experience they expect, editor and Run button and hidden tests on submit, and the reviewer gets something deliberately different. A test result arrives as a citable evidence row and is never mapped onto a rubric dimension, because a reviewer who sees a score stops reading the rubric, and a product that decides for them has quietly become an auto-reject filter.

  • Expected outputs never enter the sandbox
  • No test result ever reaches a rubric dimension
  • The agent proposes a scorecard; a TOTP commits it
  • An empty recording policy is invalid, not permissive
dossier 58
14 / 16Web + iOS/Android· Independent · Next.js · Expo · Supabase

Iron & Chalk

The maths a lifter actually needs.

A strength-training platform on web and mobile, built on a headless engine package the two clients share — estimated 1-rep max, autoregulation, readiness, load and strength standards, every number under test rather than trusted.

  • One tested engine package, no duplicated maths
  • One engine, two clients: web and Expo
  • e1RM · RPE · autoregulation · deloads
  • Research-led, not a hunch
ironchalk
15 / 16CAD pipeline· Independent · FastAPI · build123d

Jewelry CAD

A photo in, a castable part out.

A photo → castable CAD pipeline for jewellery: quality gate, VLM analysis, reference-object measurement, template retrieval, component-based parametric CAD, STEP/STL export and manufacturability validation.

  • VLM emits a typed DesignSpec
  • Deterministic code emits the geometry
  • A model error can't make an uncastable part
  • 22-document review incl. patent prior art
dossier 54
16 / 16Pricing engine· Escape Plan · FastAPI · pure pricing core

Checkmate India

The AI is allowed to fail. The price still works.

A trip-planning pricing engine where Google Places supplies the real-world facts and a 23-number rate card supplies costs, combined by deterministic code into one reproducible per-person price with margin and GST.

  • LLM deliberately kept out of the price path
  • Geographic base consolidation + room bin-packing
  • Rate-card-versioned cache keys
  • Tested and rejected Places pricing, documented why
dossier 34
Now

Currently building

Currently building: agent infrastructure at DevX AI Labs

Giving every employee a Claude-native door into the platform.

Pulse MCP is a remote Model Context Protocol server exposing a delivery-governance platform through 35 RBAC-scoped tools across 13 domains. It ships with a custom OAuth 2.1 Authorization Server — Dynamic Client Registration, PKCE, RS256/JWKS, Google Workspace federation — written because MCP clients require DCR and no managed provider offered it. The MCP layer is a stateless identity broker with zero authority to decide access: it owns no database and no domain rules, and the product re-derives role and project scope on every call.

01Authenticate
02Broker
03Audit
35RBAC-scoped tools across 13 domains
OAuth 2.1Authorization Server, written from scratch
0authority the MCP layer holds over access
dossier 16

Built on the same platform

  • Emissary Google ADK · Vertex AI Agent Engine
  • Hiring Assistant Discovery Engine · evidence reranking
  • Hiring MCP Gateway AWS Lambda · zero data at rest
  • Brand-POC Factory Shopify → grounded assistant in a day
  • Pulse MCP Facade per-tool RBAC · full audit trail
  • Atlas · Pulse the platforms the agents answer for

More work

Also shipped.

Libraries, tools and independent projects — the smaller pieces, with the detail that made each one worth building.

TradingLogic Packages

Six headless packages with zero runtime dependencies, shared by two production products under one rule: share behaviour, never share appearance. They were extracted after a duplicated implementation summed a real book of 14 wins and 10 losses to +59.87R — planned R:R counted where achieved R was meant. The regression test is named after that bug.

LIBRARY

VaghaniX

A zero-runtime-dependency macOS CLI managing multiple authorised Claude accounts — Keychain storage, atomic credential switching under file locking, a loopback-only failover proxy, and a doctor command that diagnoses without printing secrets.

CLI

Reel → Itinerary

A saved travel reel becomes three fully costed day-by-day trips. Pulled 11 real destinations from a video whose transcript was Hindi-only, via a transcript → description → caption → hashtag fallback chain.

WEB

EverTrustJewels

A diamond and jewellery storefront whose configurators keep their entire state in the URL, so any stone and setting a customer lands on is already a shareable link.

WEB

Gopinath Diamonds

A hand-modelled faceted round-brilliant diamond in WebGL with a custom refraction material and normal-map capture — five colour themes, working touch rotation, shipped end to end in three days.

WEBGL

Dukaan → twt

Reconciled 410,609 orders across all 32 source fields. Eliminated two whole false-positive classes before the client saw them — 7,385 flagged coupons became 1,231 genuinely missing, and an apparent loyalty data loss was a field mapping error with 0 real discrepancies.

DATA

Brand-POC Factory

Any Shopify storefront becomes a grounded AI shopping assistant in a day, from one declarative YAML per brand. Deployed for three. Per-turn latency cut from ~12s to 4–5s.

AI

Lead-gen System

A pipeline across nine independent discovery sources with tiered owner lookup, email verification and a built-in CRM — where consent, opt-in gating and suppression are enforced by the pipeline itself rather than left to whoever runs it.

PYTHON

About

Meet Vaghani
Meet VaghaniGCP ACE
SURAT, INDIA

Two years of building at the intersection of AI platforms, data at scale and production systems. I write the agent infrastructure, the pipelines underneath it, and the interface on top.

Currently SDE-1 at DevX AI Labs, where I authored a remote MCP server with a custom OAuth 2.1 Authorization Server, built the hiring and delivery-governance platforms it sits over, wrote the two agents that answer for them, and migrated 35 TB across 5.6M+ files with 100% verified integrity and zero errors.

Alongside that, the complete AI layer of a live trading platform at GlideTech and a hiring-assessment platform built so a machine assembles evidence and a person decides — plus two independent production security findings, and a luggage comparison platform scoring 4,500 products against 15,291 analysed reviews.

  • B.TECH CSE · GPA 8.26
  • GOOGLE CLOUD ACE
  • 2 SECURITY FINDINGS
  • MCP · ADK · VERTEX AI

Now playing

How the work got here.

Nine scenes, from coursework in C to an OAuth 2.1 server written because nobody sold one. Under thirty seconds.

How the work got here00:00:00:00
A Meet Vaghani story

How the work got here

Every number in it is on the record.
01/09

Three acts, nine scenes

Read it instead — same film, with the numbers left in.

Act I2021—2024

Practice.

Four years of theory, two internships inside a single summer, and a first job spent underneath the product rather than on top of it.

Scenes 01—03
  1. 01
    2021—25Pandit Deendayal Energy University
    Four years of theory.

    B.Tech in Computer Science, 2021 to 2025, out at 8.26. The public trail starts in May 2024, when the coursework went up in one batch — cryptography algorithms in C, a hostel manager in Python, a macromolecular classifier in a notebook.

    None of it was a product. All of it was practice.

  2. 02
    May—Jul 2024Woosong University · Superior Exports
    Two internships, one summer.

    May to July 2024 ran twice over. One half was research: AI/ML for smart buildings at Woosong University in South Korea, predictive models in TensorFlow at 85% accuracy, energy optimisation validated at 25% across ten-plus building systems, written up as a paper. The other half was delivery: a MERN B2B platform for the diamond trade at Superior Exports, 40% faster after query tuning, lazy loading and caching.

  3. 03
    Dec 2024AppArrow Infotech
    The first job was infrastructure.

    Terraform, Python and Bash on GCP — automated deployments, CI/CD pipelines, IAM under least privilege. Build errors down 15%, build times down 25%. Alongside it, a multi-tenant School ERP in Next.js and PostgreSQL with custom subdomain routing and role-based auth. A year spent underneath the product before building on top of one.

Act II2024—2026

Scale.

The title still said infrastructure. The work had already become 35 terabytes, agent infrastructure, and an authorization server that had to be written because nobody sold one.

Scenes 04—06
  1. 04
    The turnInfrastructure → product
    Then the work changed shape.

    The resume still said GCP Developer. The work had already stopped being infrastructure and become product engineering — backend, data pipeline, frontend, and the agent layer sitting above all three. The title was running about a year behind what was actually shipping. Closing that gap is most of what the last year has been.

  2. 05
    2026DevX AI Labs · Music Tribe
    Thirty-five terabytes.

    35 TB across 5.6 million files moved from Alibaba Cloud OSS to Google Cloud Storage. Zero downtime, a 100% verification match, no errors, and deduplication reclaiming 8.85 TB at zero egress cost. The finding that mattered was counter-intuitive: 50 concurrent workers failed 26 of 58 transfers where 20 workers failed 5 of 38.

    More parallelism was worse — measured, not assumed.

  3. 06
    2026MCP server · Emissary
    Writing the auth server nobody sold.

    Pulse MCP exposes a delivery-governance platform through 35 RBAC-scoped tools across 13 domains. MCP clients require Dynamic Client Registration and no managed provider offered it, so the OAuth 2.1 Authorization Server was written from scratch — RFC 7591, PKCE, RS256/JWKS, Google Workspace federation. Above it sits an ADK agent built on the same refusal to trust a model with identity: it never states who it is, and the product re-derives role and project scope on every single call. It started as a router with three specialists, until tracing showed the handoff cost 2.8 seconds a session and bought nothing, so it collapsed to one agent.

    The layer that brokers access is allowed to decide nothing.

Act III2026—

Everything at once.

Three engagements on one calendar, two security findings nobody asked for, and a way of building that had turned up in four projects before it had a name.

Scenes 07—09
  1. 07
    Jun—Aug 2026CarryScore · GlideTech
    Three engagements at once.

    June added tech lead on CarryScore — around 4,500 products scored 0–100 against 15,291 analysed reviews, gated so a score is suppressed entirely below ten. August added the whole AI subsystem of a live trading platform at GlideTech, with an LLM compliance classifier standing over every model output. TradingLogic AI went from nothing to a shipped product inside one week, sole author.

    Three engagements, one calendar.

  2. 08
    Sep 2026TradeScanner Pro
    The week the malware showed up.

    On 11 September a routine merge from origin pulled in a payload hidden in next.config.js — 41 injecting commits, all 54 remote branches, live for about eleven weeks. Its string decoder was reimplemented in Python so the payload could be read without ever being run. Eleven days earlier came a second, unrelated finding: a JWT bypass across all eight WebSocket endpoints.

    Both were written up. Neither was the job.

  3. 09
    TodaySurat, India
    Today.

    Around half a million lines across the projects, sole author on eight of them, production systems on four clouds, two independent security findings, thirty-odd written deliverables. Underneath all of it sits one idea that turned up independently in four separate projects before it had a name.

    The model decides what. The code decides how. And a system refuses before it guesses.

Experience

Where I've built.

From a multi-tenant School ERP to agent infrastructure, a 35 TB migration and the AI layer of a live trading platform.

6 ROLESSINCE MAY 2024
  1. JAN 2026 — PRESENT

    DevX AI Labs

    INDIA
    01 / 06NOW

    SDE-1

    • Built Pulse MCP — a remote Model Context Protocol server exposing 35 RBAC-scoped tools across 13 domains, with a custom OAuth 2.1 Authorization Server (RFC 7591 DCR, PKCE, RS256/JWKS, Google Workspace federation) written because no managed provider offered DCR
    • Architected a two-layer auth model — OAuth for users, short-lived HMAC identity tokens service-to-service — with the MCP layer as a stateless identity broker that owns no database and no domain rules
    • Migrated 35 TB across 5.6M+ files between Alibaba Cloud OSS and GCS with a distributed pipeline — zero downtime, 100% verification match, 0 errors, ~60% deduplication reducing storage by 8.85 TB at zero egress cost
    • Built a SKU-level demand-forecasting engine reaching 14% monthly WAPE against a 27% naive baseline, powering a ₹3.2 Cr / 63k-unit monthly plan across 11 channels
    • Owned reconciliation for a 410,609-order commerce migration across all 32 source fields — 92.1% clean, and eliminated two entire false-positive defect classes before they reached the client
    • Built two production agents on Google ADK + Gemini over Vertex AI Agent Engine — a project-governance agent doing ref-based reads and writes behind an HMAC callback token, and a recruiter-facing candidate assistant with evidence-based reranking on Discovery Engine
    • Built the hiring platform behind that assistant — candidate journey from application through screening, scheduling, rounds, offer, onboarding and probation, with self-booked scheduling links, Google Calendar watch channels and Meet artifacts over Pub/Sub, and interview feedback drafted from the transcript on AWS Bedrock
    • Built the delivery-governance platform — RAID and action registers, project plan, weekly reviews scored against a weighted health matrix, CSAT campaigns and a client portal — with a seven-signal matcher filing every calendar meeting against the right project, and a calibration rule that a cosine score may surface a suggestion but only a reranker score may act on one
    • Ran release and merge management across 530+ pull requests and 23 contributors
  2. AUG 2026 — PRESENT

    GlideTech Solutions

    REMOTE
    02 / 06NOW

    Full-Stack & AI Engineer

    • Built the entire AI subsystem of a live retail-trading platform — signal explainer, trade-plan builder, streaming conversational coach with tool-calling, document and vision analysis, a pluggable RAG stack and a trading journal, ~60 modules behind a two-tier cost-routed model layer
    • Implemented the product's regulatory boundary as runtime code — an LLM compliance classifier over every model output blocking personalised advice and price predictions, and detecting jailbreaks and prompt injection
    • Made the AI refuse instead of hallucinate: added grounded/asOf provenance to every generated plan and an explicit noData refusal when market data is missing
    • Cut equity-chart response time from 26s cold to 1.5s (17×) and 1.35s warm to 9ms (150×), and migrated the live ticker from REST polling to WebSocket
    • Sole author of TradingLogic AI — 112 commits, 13 routes, ~90 components, ~18,400 LOC in one week, with a 22-primitive design system, a full GSAP motion system and React Three Fiber scenes
    • Sole author of a six-package headless shared-logic library (~9,800 LOC, zero runtime dependencies) consumed by two production products, with the rule 'share behaviour, never share appearance' enforced as runnable assertions
    • Discovered active supply-chain malware in a production repository and reverse-engineered it statically — reimplementing its string decoder in Python so the payload was never run — then shipped a config-integrity gate testing shape rather than signatures
    • Sole author of ProofHire, a multi-tenant hiring-assessment platform across three services — FastAPI backend, Next.js client and an MCP server — built so a machine assembles evidence and a person decides: a coding-stage test result is a citable evidence row and is never mapped onto a rubric dimension
    • Designed its execution sandbox so the expected outputs never enter it — the runner returns what the program printed and the API grades, so a sandbox escape cannot exfiltrate hidden test answers — and gated agent-written scorecards behind a TOTP commit bound to a hash of the proposed payload
  3. JUN 2026 — PRESENT

    Escape Plan / CarryScore

    CLIENT ENGAGEMENT
    03 / 06NOW

    Tech Lead

    Delivered through DevX AI Labs

    • Tech lead on CarryScore, a live luggage comparison platform indexing ~4,500 products across 9 brands and 20+ specs — FastAPI + PostgreSQL 16 on AWS, architected so the app boots with no database while live prices hot-reload with zero restarts
    • Designed the CarryScore 0–100 rating algorithm — five dimensions weighted across six buyer profiles, blended 50/50 against an AI review-analysis pipeline over 15,291 Amazon and Flipkart reviews, gated by confidence tiers that suppress a score entirely below 10 reviews
    • Ran a technical SEO audit across 4,518 crawled pages and traced ~4,495 unrankable pages and a 2.4 MB per-view payload to one root cause, then shipped the remediation as a Next.js 16 SSR/ISR migration
    • Built the Checkmate India trip-pricing engine (FastAPI) with the LLM deliberately kept out of the pricing path, so the AI itinerary writer is allowed to fail and the wizard still returns a price
    • Turned a destination management company's 2025/26 B2B rate card — 15 raw Excel/PDF files, 658 seasonal rate rows — into 12 normalised datasets and a 674-line pricing engine designed to refuse rather than guess
  4. DEC 2024 — JAN 2026

    AppArrow Infotech

    INDIA
    04 / 06

    Full Stack Developer

    • Developed a multi-tenant School ERP system using Next.js, TypeScript and PostgreSQL, integrating attendance, fees, transport and communication modules
    • Implemented role-based authentication, custom domain/subdomain routing and modular workflows for security and scalability
    • Conducted requirement analysis, system design and internal testing
  5. MAY 2024 — JUL 2024

    Woosong University

    SOUTH KOREA (REMOTE)
    05 / 06

    Research Intern

    • Researched AI/ML-driven smart building systems, designing predictive models with Python and TensorFlow for IoT-enabled automation at 85% prediction accuracy
    • Proposed and validated energy-optimisation algorithms improving efficiency 25% across 10+ building systems, and authored a research paper on predictive maintenance and intelligent automation
  6. MAY 2024 — JUL 2024

    Superior Exports

    INDIA
    06 / 06

    Full Stack Developer Intern

    • Built a full-stack B2B e-commerce platform for the diamond industry on the MERN stack, improving website performance 40% through code optimisation, lazy loading, caching and MongoDB query tuning

Stack · tech radar

The stack behind the work.

36 tools across 6 groups, plotted by proficiency: the closer to the core, the stronger. Hover or tap a group to light up its sector.

MCP — server, gateway, facadeGoogle ADKVertex AI Agent EngineRetrieval & rerankingAWS BedrockAgent safety & evals36TOOLS

CORE = STRONGEST · RINGS AT 90 / 80 / 70

AI & Agents

6 on the radar
  • MCP — server, gateway, facade96%
  • Google ADK92%
  • Vertex AI Agent Engine90%
  • Retrieval & reranking88%
  • AWS Bedrock86%
  • Agent safety & evals85%

Also in the toolbox

  • MySQL
  • MongoDB
  • Firestore
  • Cloud SQL
  • DynamoDB
  • Supabase
  • OpenSearch
  • Discovery Engine
  • Cloud Pub/Sub
  • Cloud Run
  • Cloud Functions
  • Cloud Build
  • Google Calendar API
  • Google Meet API
  • Google Drive API
  • AWS Lambda
  • S3
  • EKS
  • WebSockets
  • SSE
  • A2A
  • Playwright
  • Vitest · pytest
  • Medusa.js
  • Strapi
  • Shopify Liquid
  • Three.js
  • Zustand
  • Radix
  • Astro
  • scikit-learn
  • XGBoost
  • TensorFlow
  • Langfuse
  • GitHub Actions
  • Vercel
  • Slack API
  • Nodemailer

Speaking globally

000+ talks. 00+ cities. 00 countries.

One line on where you speak and what you speak about, from local meetups to international conferences.

00Countries
00+Cities
000+Talks & workshops

Arcs from home

Your City to the world

YOUR CITY · HOME

Closer to home

  • City 01
  • City 02
  • City 03
  • City 04
  • City 05
  • City 06
  • City 07
  • City 08
  • City 09
  • City 10
  • City 11
  • City 12
  • City 13
  • City 14
  • City 15
  • City 16
  • City 17
  • City 18
  • City 19
  • City 20

Let's build together

What's next? Let's make it together.

Work with me

Have a system to build?

AI agent infrastructure, backend and data platforms, and the full-stack product on top — from architecture to production.

Talk to me

Want to go deeper on any of this?

Happy to walk through any project above — the architecture, the trade-offs, or the things that didn't work.

Send an email—· usually reply within a day