(Portfolio ©2026)

Senior Software Engineer
AI systems · Backend · Full-stack

Currently at LTIMindtree
Based in India

I build AI products that show their work — the model makes the calls, tested code does the math, and a human approves anything that can't be undone.

I'm a software engineer from IIT Jodhpur who builds backend systems and AI products. I care less about what a model can say and more about what it can prove — so language models make the decisions, and the numbers live in code that has tests.

By day I build backend services at LTIMindtree. Before that, GenAI tooling for clinicians at Wavo Health. In between, I ship products of my own — two of them are below.

  • 10K+internal users on backend services I built at LTIMindtree
  • 25%faster average API responses after query, index and cache work
  • 40%less clinical documentation time with AI note automation
  • <2sresponse latency on a WhatsApp + web AI support agent
PythonTypeScriptNext.jsFastAPILangGraphClaude APIPostgreSQLReactThree.jsFlaskRAGGeminiDockerWebGLExpoNode.js
PythonTypeScriptNext.jsFastAPILangGraphClaude APIPostgreSQLReactThree.jsFlaskRAGGeminiDockerWebGLExpoNode.js

Nine things I designed and built in 2026. Five put a language model inside hard guarantees and then show the user exactly what it did; four are real-time WebGL, where the whole film runs off your scroll position.

  1. 01 Parley — a chatbot your site can own
  2. 02 Jobflow — a job hunt on autopilot
  3. 03 Nightshift — an AI on-call engineer
  4. 04 Dossier — multi-agent research, cited
  5. 05 Payoff — debt, planned by an agent
  6. 06 Proof — receipts, deadlines, reminders
  7. 07 Helixa — 3D medicine at every scale
  8. 08 Aranya — 3D flat tour
  9. 09 RONIN — steel that becomes ink
  10. 10 HADAL — a dive to the Challenger Deep
(01)AI · SaaS · Multi-tenant2026

Parley

Your website, answering for itself.

Point it at your site

It follows the links and the sitemap, splits what it finds into passages and embeds them. Nine pages became twenty-seven passages here, re-crawled daily.

It answers, with receipts

Every answer carries the passages it came from. With no AI key it quotes them outright; with one it writes the reply and still cites. It only answers from what you gave it.

Embed it anywhere

One script tag, or React, WordPress, an iframe, a CDN build or the JS API — all the same assistant. Allowed domains, not a secret, are what protect the key.

parley · knowledgeRunning locally
Point it at your siteIt answers, with receiptsEmbed it anywhere
Role
Product, design & engineering
Year
2026
Type
AI · SaaS · Multi-tenant

A chatbot anyone can put on their own site: point it at your pages or files, and it crawls, chunks and embeds them into pgvector. Visitors get answers with the passages they came from, and you get the install snippet — script tag, React, WordPress, iframe or the JS API.

What makes it a product, not a demo

Every tenant's answers stay inside their own pages.

Each account gets its own bot, its own sources and its own embeddings; retrieval is scoped to them, so one customer's assistant can never quote another's site. The widget renders in a shadow root, which is why it can't inherit — or leak into — the host page's CSS. Public bot keys are safe to ship in HTML because allowed domains, not the key, are what authorise a site.

index.html
<!-- anywhere before </body> --><script  src="https://cdn.parley.app/widget.js"  data-bot="pk_live_aurora_dental"></script>// or, in a React app:import { Parley } from "parley-react";<Parley botKey="pk_live_aurora_dental" />
  • 102tests, including Postgres integration
  • 6ways to embed one assistant
  • 2SDKs published to npm with provenance
  • Express 5
  • Drizzle
  • Postgres + pgvector
  • Preact
  • Shadow DOM
  • Next.js 16
  • npm packages
  • Docker
(02)AI · Agents · Automation2026

Jobflow

A job hunt on autopilot, with a human at the wheel.

01 / 06

Discover

Seven job boards queried in parallel, deduplicated, and checked for freshness before anything is stored.

Match

Every posting scored 0–100 against your CV — and every point comes with the evidence behind it.

Tailor

Claude or Gemini rewrites the CV per job. Inventing employers, dates or skills is forbidden, and gaps are surfaced.

Find a contact

Only published addresses — the posting, the company's own site, or Hunter.io. Never guessed from name patterns.

Your review

The graph pauses. Nothing leaves your Gmail until you've read each draft and ticked it.

Send & summarise

Only what you approved goes out, within the daily limit. Every step was checkpointed, so a crash resumes cleanly.

jobflow · Autopilot · run #14Illustrative
  1. Discoverdiscover_jobs
  2. Matchselect_jobs
  3. Tailortailor
  4. Find a contactfind_contacts
  5. Your reviewreview · interrupt()
  6. Send & summarisesend_approved

Sources · fetched this run

  • Adzuna112
  • JSearch89
  • Himalayas61
  • Arbeitnow47
  • Remote OK33
  • Remotive24
  • Jobicy18

384 fetched · 57 new · 12 skipped as expired

Top matches · score ≥ 75

  • 91Senior Backend EngineerLumen Labs · RemoteHimalayas
  • 86AI Engineer, AgentsNorthstar AI · BengaluruAdzuna
  • 78Full-Stack Engineer, PythonKitebase · RemoteRemotive

Skills 47/55 · Role 18/20 · Location 15/15 · Salary 7/10

What changed · Senior Backend Engineer

  • ↑Moved distributed systems to the top of skills — the posting names it 4×
  • ✎Led the LTIMindtree bullet with the 25% API latency win
  • +Surfaced the RAG marketplace project for “LLM retrieval”
  • !Kubernetes is required and not in your CV — flagged, not added

CV-Satender-Maurya-Lumen-Labs.pdf · cover letter drafted

Hiring contacts · lumenlabs.io

Found on lumenlabs.io/careers · robots.txt respectedPublished
Hunter.io · 94% confidence · source linkedHunter

0 addresses guessed · 0 mail servers probed

Paused — waiting for you

Tocareers@lumenlabs.io
SubjectSenior Backend Engineer — Satender Maurya

Hi — I build backend services and AI tooling, most recently cutting API latency 25% at LTIMindtree. I've attached a CV tailored to the role…

Send this oneApprove & send 2

Run complete

  • 3applications prepared
  • 2emails approved & sent
  • 1kept as a draft
  • 0failures

Checkpointed to workflows.db · daily limit 2 / 20

Role
Product, design & engineering
Year
2026
Type
AI · Agents · Automation

Jobflow finds jobs across seven boards, scores each one against your CV with the evidence behind every point, rewrites your CV and cover letter for the ones you pick, finds a published hiring contact, and drafts the email. A LangGraph autopilot runs the whole loop — then stops and waits for you before anything is sent.

Explainable match score

A number you can argue with.

The matcher is deterministic and needs no API key. It returns the evidence behind every point, and editing your profile rescores the whole board.

  • Skill overlap47/55
  • Role fit18/20
  • Location15/15
  • Salary7/10

Guardrails

  • 01Nothing sends without a dialog showing the exact recipient, subject, body and attachments.
  • 02Contacts come only from published addresses — never guessed from name patterns.
  • 03At most 20 emails a day, and no repeat to the same person about the same job within 24h.
  • 04Every graph step is checkpointed, so an interrupted run resumes without duplicating work.

Sources

  • Adzuna
  • JSearch
  • Remotive
  • Himalayas
  • Jobicy
  • Remote OK
  • Arbeitnow
  • 7job sources, deduplicated and freshness-checked
  • 3tailoring engines: Claude, Gemini, or rule-based
  • 0emails sent without your approval
  • Flask
  • SQLAlchemy
  • LangGraph
  • Claude + Gemini
  • Next.js 16
  • TypeScript
  • Gmail OAuth
  • ReportLab

Three more, built end to end

(03)AI · Agents · SRE2026

Nightshift

An AI on-call engineer that tries to prove itself wrong.

When an alert fires, Nightshift plans an investigation and sends specialist agents through metrics, logs, deploys, feature flags and runbooks in parallel. It tries to disprove its own hypotheses, then hands the on-call human a root cause where every claim links to the evidence behind it.

  • 01Every model turn, tool call and approval is a durable step. Kill a worker mid-incident and another resumes from the last finished step — no repeated LLM calls, no restarted deploys.
  • 02Fixes are typed actions behind role and blast-radius policy: they wait for a human, execute durably, then check whether the alert metric actually recovered.
  • 03With no API key it runs deterministic baseline agents — same tools, same orchestrator, same guardrails, at zero cost.
  • 272tests, run against SQLite and Postgres
  • 20simulated incidents scored against ground truth
  • 6integrations, emulated end to end before anything real
  • FastAPI
  • Claude API
  • Postgres
  • MCP
  • Next.js
  • Playwright
  • Docker
  • Stripe
nightshift · incident · payments-pool-config
Nightshift incident dashboard: a diagnosed checkout error-rate alert with planner, parallel investigators, hypotheses and verifiers
A real run — baseline agents, no API key
(04)AI · Multi-agent · Research2026

Dossier

A report that has already been criticised — by the system that wrote it.

Give Dossier a topic and it plans sub-questions, runs a researcher per sub-question in one parallel step across the web and your own documents, writes a cited draft, fact-checks it, criticises it, and revises. Every claim cites a numbered source; the model never writes a URL, and searches that fail are reported as gaps instead of filled in from memory.

  • 01Upload a PDF and a second graph turns it into a course: lessons in the order the ideas have to be learned, each explained as if to a five-year-old and then properly.
  • 02Four ways in — a CLI, an HTTP API with live progress, an MCP server for Claude Desktop, and a Next.js front end.
  • 03The evals are honest about themselves: the revision loop never fired, so it's documented as insurance rather than an improvement path, with a measured judge noise floor of 0.04.
  • 27topics judged by a model other than the writer
  • 0invented citations across every report
  • 245tests, all runnable without an API key
  • LangGraph
  • Gemini
  • Tavily
  • Chroma
  • FastAPI
  • SSE
  • MCP
  • Next.js
dossier · research graph
topicvalidateplannerweb · sub-question 1web · sub-question 2docs · sub-question 3docs · sub-question 4writerfact-checkcriticfinalizeone parallel steprejected · up to 3 drafts

Judged by a different model · 27 topics

  • Groundedness4.85
  • Completeness4.74
  • Coherence5.00
  • Usefulness4.74

0 invented citations · 100% passed review first draft

(05)AI · Fintech2026

Payoff

Every loan is a gravity well.

Upload a bank statement. Payoff finds the loan payments, works out what you can actually afford, and an agent builds a payoff plan — with every number traced back to a tested amortization engine instead of a language model's guess.

  • 01The agent's tool schemas accept scenarios — an extra payment, avalanche or snowball — but never balances, rates or EMIs. Those live in a closure it cannot write to, so a fabricated figure is structurally impossible rather than merely discouraged.
  • 02CSV and PDF readers infer the columns, the date order and the sign of every amount, then recurring-payment detection separates EMIs from subscriptions and measures real monthly surplus.
  • 03Statements are parsed in memory and never written to disk. A store of customers' transactions is a breach waiting to happen; not having one is the feature.
  • 22tests on the engine, parser and detection
  • 0figures computed by the language model
  • 0bank statements written to disk
  • Next.js 16
  • TypeScript
  • Claude API
  • Tool use
  • Zod
  • WebGL
  • unpdf
  • Node test runner
The hero shader from Payoff — here the scroll pays the loan down
(06)Mobile · Consumer2026

Proof

The deadline you forgot you had.

Snap a receipt and Proof tracks every deadline that matters: the return window, the manufacturer warranty, and the EU legal guarantee most people never claim. It reminds you before each one runs out.

  • 01No backend and no account. Purchases live in SQLite on the phone; only the receipt photo leaves the device, to be read by Google Cloud Vision.
  • 02An offline parser pulls shop, date, total, return days and warranty length out of receipts in nine European languages.
  • 03Reminders are rebuilt after every change and capped at 60, because iOS only allows 64 pending notifications.
  • 35European retailers' return policies built in
  • 9receipt languages parsed offline
  • 2app stores from one Expo codebase
  • Expo
  • React Native
  • expo-router
  • SQLite
  • Cloud Vision OCR
  • TypeScript
ProofIllustrative

Closing soon

  • Wireless headphones3 days leftAmazon · delivered 12 SepReturn by 12 Oct
  • LaptopEU guarantee · 2 yrsApple · bought 4 MarWarranty to Mar 2028
  • Running shoes62 days leftZalando · 100-day policyReturn by 21 Dec

12 reminders scheduled · nothing leaves the phone but the receipt photo

Four sites where scroll is the only input

Real-time WebGL, written the same way each time: one scroll position feeds one state object, and that object decides the camera, the light and the copy. Nothing here is a video, and almost nothing is downloaded — the geometry is generated in code.

Helixa: a DNA double helix rendered in 3D beside the headline 'Read the code. Then rewrite it.'(07)Live

HelixaMedicine, at every scale

One continuous descent from molecule to organ for a precision-medicine brand. Each chapter pins a live specimen that turns, cuts itself open and labels its parts as you scroll.

The heart and brain are signed distance fields polygonised with marching cubes in a Web Worker, with ambient occlusion and anatomical regions baked per vertex. Nothing is downloaded — every model is generated in code.

  • Genomics
  • Cell biology
  • Immunology
  • Delivery
  • Cardiology
  • Neurology

Next.js · react-three-fiber · GSAP · Web Workers

helixa.sattu.world
Aranya: 3D residential towers above a colourful city, with hotspots labelling a residence and sky lounge(08)

AranyaA flat tour, flown not clicked

A drone shot over the towers, a flight into one flat, then room by room to the amenities — the whole tour on a single scroll, with callouts that arrive as the camera does.

Scroll maps to a Catmull-Rom camera path through twelve stops. Geometry is procedural and merged, dressed with CC0 PBR materials and image-based lighting.

  • Drone
  • Tower
  • Living
  • Kitchen
  • Bath
  • Amenities

Next.js · react-three-fiber · drei · Post-processing

RONIN: a black ink silhouette of a samurai in a bamboo grove under rain and a red sun, on aged paper(09)

RONINSteel that becomes ink

Eleven chapters of a samurai story. A katana is photographed in steel on four-hundred-year-old paper, then becomes sumi ink on that same paper as the story turns.

Steel, ink, rain, petals, birds, the sun and the paper are all generated — there is not one photograph, texture or model in the repository. Every timeline is keyed in viewport units, so a chapter boundary is the same number everywhere.

  • The seal
  • The forging
  • The oath
  • The grove
  • The duel
  • The stillness

Next.js · react-three-fiber · Custom shaders · Lenis

HADAL: a submersible lit by floodlights in dark blue water beside the chapter title 'The Twilight Zone'(10)

HADAL10,935 metres down

A descent to the floor of the Challenger Deep aboard a procedurally modelled submersible. Scroll position is the only input to the whole film.

One state object turns depth into water colour, fog, lamp power, seafloor height, trench width and fifteen camera keyframes, once per frame. The pressure sphere shrinks by 0.7% of its diameter on the way down — about what 1,100 bar really does to it.

  • Surface
  • Twilight
  • Midnight
  • Abyssal
  • Trench
  • The Deep

Next.js · react-three-fiber · Instancing · Lenis

Frames captured from each site running at full frame rate, not mockups.

How I build

Four rules that came out of shipping Payoff and Jobflow. They're the reason people can trust software that has a language model inside it.

01Seen in Payoff

Models decide. Code computes.

Payoff's agent can choose which scenarios to run, but its tool schemas don't accept balances or rates. A made-up figure isn't discouraged — it's structurally impossible.

02Seen in Jobflow

Irreversible means approved.

Jobflow's autopilot drafts everything, then the graph pauses. Only the emails you tick leave your Gmail — and it can send as you, but never read your inbox.

03Seen in Payoff

Don't keep what you can't protect.

Bank statements are parsed in memory and discarded. A table of customers' transactions is a breach waiting to happen; not having one is the feature.

04Seen in Dossier

Measure it, then believe it.

Dossier's reports are judged by a different model from the one that wrote them, and the noise floor between two judgements of the same text is published — so any 'improvement' smaller than 0.04 is known to be noise.

05Seen in Both

Say the limits out loud.

Fixed-rate assumptions, currency-blind salary matching, text-layer PDFs only — stated next to the result, not buried in a README.

Where I've shipped

  1. Jul 2025 — Now

    LTIMindtree

    Senior Software Developer · India

    • Backend services used by 10K+ internal users across business units, built for high availability.
    • Secure REST APIs with JWT authentication and role-based access control.
    • SQL, indexing and caching work that cut average API response time by 25%.
    • Shipped 5+ production features alongside frontend, QA and DevOps teams.
  2. Jun 2024 — Sep 2024

    Wavo Health

    Software Developer · Remote

    • AI note automation for healthcare workflows — documentation time down 40%.
    • A GenAI assistant on LangChain and OpenAI for real-time patient summary retrieval.
    • Storage and indexing pipelines for 1K+ patient visit summaries.
  3. 2021 — 2025

    IIT Jodhpur

    B.Tech, Mechanical Engineering · Rajasthan

    • Data structures, algorithms and object-oriented design, alongside an engineering degree.

Toolkit

Languages

  • Python
  • TypeScript
  • JavaScript
  • C / C++
  • SQL

Backend

  • FastAPI
  • Flask
  • Django
  • Node.js
  • REST design
  • JWT / RBAC

Frontend

  • React
  • Next.js
  • Tailwind CSS
  • GSAP
  • Expo

3D / graphics

  • Three.js
  • react-three-fiber
  • GLSL shaders
  • WebGL
  • Canvas

AI

  • Claude API
  • Gemini
  • LangGraph
  • LangChain
  • Tool use
  • RAG
  • Evals
  • MCP

Data & infra

  • PostgreSQL
  • MongoDB
  • SQLite
  • Docker
  • Playwright
  • pytest

Have an idea?Let's build it.

Email me

Open to conversations about AI products, backend systems and full-stack roles. I usually reply within a day.