Yash Kumar, AI Application & Full-Stack Engineer, Résumé

Summary

AI application and full-stack engineer, four years shipping and then operating what I build. I run a live LLM chatbot in production — a multilingual clinical assistant on a client’s website, answering patients in Hindi, English or Hinglish and steering them into the booking flow — and a retrieval-augmented knowledge-base service with a document ingestion pipeline, pgvector search, cross-encoder reranking and an evaluation harness that gates CI. I integrate LLM APIs across Anthropic, OpenAI, Google and DeepSeek, and build the web applications and REST APIs around them: 39 services and 17 production databases on a single VPS, all mine end to end.

39services in production

17PostgreSQL databases

1,572API routes across six products

527tables in those six schemas

142SQL migrations authored

23live hostnames, one VPS

Counted from the live estate on 13 September 2026 by a script kept beside this page’s source, with third-party code excluded. Every system named below is running now and can be demonstrated end to end.

Measured on the live estate, 13 September 2026: 39 services in production, 17 PostgreSQL databases, 1,572 API routes and 527 tables across six products, 142 SQL migrations, 23 live hostnames on one VPS. Counted by a script kept with this résumé’s source, third-party code excluded. Every system named below is running now and can be demonstrated end to end.

AI & Chatbot Engineering

“Dr Anshuman” clinical assistant · production LLM chatbot on a client website geetasriretina.org

Shipped a multilingual patient-facing chatbot that is live on a real clinic’s website, answering in Hindi, English or Hinglish — it mirrors whatever language the patient types in, and three UI pills can force a language, implemented as server-side prompt overrides appended after the base rule so the override wins even when the patient writes in another language.
Built the whole path myself: a Node/Express proxy so the API key never reaches the browser, a whitelisted-role message validator that caps history at the last 10 turns and 2,000 characters per turn, a per-IP limit read from the trusted client IP rather than a spoofable header, a 20-second abort, and a fallback on every failure branch that still answers the patient and points them at booking instead of surfacing an error.
Made the model drive the interface rather than just talk: the system prompt emits a structured [[LEARN:<key>]] tag from a fixed eight-key whitelist, which the client parses to open the matching patient-education panel — so an answer about screen fatigue also opens the eye-exercise content. Clinical safety is in the prompt contract too: general guidance only, a translated disclaimer on every medical answer, and an explicit escalation to an in-person visit on red-flag symptoms.
Closed the loop from assistant to revenue: the bot routes patients into a booking flow that issues a queue token, fires a same-day alert to the clinic’s inbox and sends the patient a confirmation with arrival instructions — replacing three notification paths that looked like they worked and reached nobody (browser push with zero subscriptions, a socket event needing the queue screen open, and a WhatsApp share link with no recipient). Reception calls back from the alert; the clinic reports roughly ₹1 lakh in additional revenue in the first 19 days after launch.
The same multi-tenant endpoint serves other clinic sites by slug, so a second clinic gets a branded assistant without new backend code.

VIXIT Cortex · retrieval-augmented knowledge base over your own documents cortex.vixitai.com

Built the full RAG pipeline end to end — PDF, DOCX and CSV parsing with OCR fallback, normalisation, multi-level chunking, batched embedding and pgvector storage — then two-stage retrieval: vector cosine search narrowed by a Voyage rerank-2.5 cross-encoder. Embeddings run on Google text-embedding-005 or Voyage, behind one provider interface so a model swap is a config change.
Designed the seven-layer query router so only two of seven layers ever call an LLM — schema, saved-query, cache and statistical lookups answer first — which keeps the common question cheap and fast instead of billing a model for it.
Made answer quality a build gate rather than a vibe: an evaluation harness with retrieval metrics, an LLM-as-judge suite and baseline comparison, wired to a CI command that fails the build on regression, with 20 dated runs kept on disk as evidence.

AI text-analysis services · Python, FastAPI, model serving

Built Python and FastAPI services serving transformers, torch and sentence-transformers models behind HTTP APIs across 24 endpoints, ensembling a fine-tuned RoBERTa classifier with a from-scratch GPT-2 perplexity and burstiness signal so a single model’s bias cannot decide the answer alone, alongside plagiarism, grammar and style analysis.

LLM operations · routing, self-hosting, day-to-day practice

Self-host and harden an LLM gateway behind TLS and three layers of auth for routing and key management across providers, and use AI-assisted development daily (Claude Code) as a normal part of shipping — including building this estate.

Experience

Founder & Full-Stack Engineer · VixitAI Labs, Ranchi

2022 to Present

Independent software work since 2022, formally registered as VixitAI Labs in 2026 (Udyam + GST). Architecture, engineering, DevOps, security and product for a multi-product estate, single-handed.

VIXIT Clinic ERP · multi-tenant clinic and hospital platform clinic.vixitai.com

Shipped 60 clinical modules across 813 API routes and a 207-table PostgreSQL schema — EMR, pharmacy POS with GST returns and a narcotics register, lab, radiology with DICOM, IPD, OT, blood bank, TPA, HR and a 56-endpoint patient portal — as the only engineer, on a two-instance PM2 cluster behind Nginx with per-tenant isolation enforced across 52 files.
Took the platform from records to revenue by building Razorpay subscription billing with live recurring mandates, GST-inclusive invoicing and module-level licensing, so a plan maps to a feature set without a code change.
Turned India’s health-data standard from a blocker into a shipped surface: 28 ABDM/ABHA endpoints covering consent, HIP callbacks and FHIR bundles, with AES-256-GCM encryption at rest for imaging.
Hardened the stack myself — least-privilege PostgreSQL roles, audit logging, JWT with token-version revocation that ends every session on password reset, a client-secret header enforced across 192+ calls, fail2ban and Nginx WAF rules at the edge.
Found a patient-identity defect by reading the writers rather than the tickets: four code paths normalised phone numbers three different ways and nothing deduplicated, so the same patient could exist three times.

Upper Bazaar · livelihood and local-commerce marketplace upperbazaar.com

Built a bilingual (English and Hindi) marketplace — 288 API routes, 104 tables, 68 forward-only migrations — covering worker profiles, shops, orders, delivery dispatch, scheduled slots, wage panels and government-scheme guidance, with a Next.js 16 / React 19 storefront and four single-file operator consoles non-technical staff run without me.
Replaced a test suite that passed silently — the repository’s configured test command matched no files at all — with 100 end-to-end scripts against an isolated stack owning its own database, Redis and ports, now permanently running rather than a folder of files.
Made the rider and shop apps work without signal: a cache-first app shell so the app opens with no network, network-only reads because a stale delivery state is a wrong instruction rather than a degraded one, a client-side action queue, and web push over VAPID to reach a rider whose screen is off.
Cut mobile Lighthouse performance from 47 to the high 90s and largest-contentful-paint from 7.4s to 2.2s through responsive image renditions and font-payload work.
Kept dispatch idempotent across deploys with a PostgreSQL advisory lock around the sweep, after finding pm2 reload could run it twice.

Visible Kitchen · kitchen operations and compliance platform

Designed and built a Fastify 5 and TypeScript platform — 112 routes, 80 tables — covering orders, ledger, tax, invoicing and settlements with operator consoles and an inspector PWA, writing the full high- and low-level design first and building to it.
Covered the money paths with 11 Vitest suites (outbox, money, ordering, RBAC, trust) and a measured phase benchmark rather than an estimate.

Jantar Mantar · short-video platform with a native Android client

Built a short-video platform — 116 routes, 43 tables, 38 migrations — with HLS transcoding, a ranking worker, creator upload, a moderation console and Cloudflare R2 storage, verified by 11 end-to-end suites carrying 416 assertions.
Shipped the Android client on Expo 54 and React Native 0.81 with over-the-air updates, building both APK and AAB on a headless VPS toolchain (JDK 17 and the Android SDK) with no hosted build service and no paid account.
Cut cold open from 2.96s to 0.29s with a playlist cache and progressive MP4 delivery.

Rentlelo · peer-to-peer rental marketplace

Took a rental marketplace from market research through a full PRD to a live service — 137 routes, 57 tables, PostGIS and pg_trgm search, Redis and BullMQ, Zod validation on every route.
Designed privacy in rather than bolted on: phone numbers never appear in list responses, only behind a rate-limited contact-reveal; identity documents are stored masked and encrypted behind signed URLs; the Play Store data-safety answers are derived from the schema rather than guessed.

Content and CMS platforms · Strapi 5 and WordPress

Run two content platforms: a Strapi 5.47 headless CMS for multi-tenant article management, and a WordPress install taken from mobile Lighthouse 65 to 95 and LCP 6.5s to 2.6s by cutting the page from 2,692KB to 764KB.

Freelance Web Developer · Independent

2021 to Present

Built and operate the patient portal for Geeta Sri Retina Institute: OTP sign-in, self-serve registration, booking, reschedule and cancel, a live queue, per-visit medical-record PDFs, automated patient and clinic email, a Hindi interface, and a multilingual assistant proxied server-side so no API key reaches the browser.
Delivered a local-SEO overhaul for the same clinic (11 articles, an indexable Hindi section, structured data), and built a client-managed photo slider on Cloudflare R2 so Ekdant Dental updates its own gallery without me.
Built and maintain a custom WordPress child theme for KoshaNews with a transient render cache invalidated on save_post, taking a cold page of 171 queries / 121ms to 0 queries / 0.2ms warm.

Selected Engineering

SPARSE · CPU-only detection of resident space objects in neuromorphic event-camera streams

Held per-window detection to 1.01ms mean and 1.46ms p99 against a 40ms real-time budget on two CPU cores with no GPU, by staying in the event domain end to end — noise filtering, O(N) grid clustering with sub-pixel centroids, then a constant-velocity Kalman filter with probabilistic data association — instead of rasterising events into frames for a CNN.
Made it reproducible by a stranger: 2,114 lines of NumPy and SciPy with 60 tests, packaged as a non-root Docker image on a read-only registry I host, so docker pull vixitai.com/sparse:1.0 reproduces the reported numbers exactly.
Found and fixed the defect that mattered most — the hot-pixel mask was deleting the stationary geostationary object the detector existed to find — and report weak accuracy on held-out real recordings rather than tuning per recording to hide it.
Rejected a YOLO-based detector on two grounds: it is the wrong tool for a two-to-five-pixel point source, and its AGPL licence is incompatible with the competition’s IP-assignment terms. Submitted to the TII OrbitSight Challenge (US$50,000); a live explore engine runs the pipeline in the browser at vixitai.com/solutions/neuromorphic/explore/.

Deedmint · on-chain registry of record for property title

Built a seven-contract Solidity registry with Foundry, a Node 22 API on raw SQL, a BullMQ settlement worker and a chain indexer, where a retried settlement batch is idempotent at three layers at once: a BullMQ job id, a PostgreSQL unique constraint, and an on-chain replay guard.
Enforced four-eyes approval twice over, in a database constraint and a 3-of-5 multisig, and kept the indexer single-instance through both PM2 and a database lease. Runs on a local Anvil development chain; nothing is deployed to testnet or mainnet.

FSF ERP · zero-dependency invoice ledger

Replaced a broken invoice form for a retail client with a ledger built on Node’s built-in node:sqlite and node:http and no npm dependencies at all, capped at 256MB under systemd with a nightly VACUUM INTO snapshot and integrity check — the smallest thing that survives unattended was the right answer.

bodemap · geospatial land-investment demo bodemap.vixitai.com/luxembourg

Built a Next.js 16 mapping demo on MapLibre with vector tiles I generate and host myself (Planetiler into tileserver-gl — Belgium 440MB, Luxembourg 7,043 tiles), including a 3D flythrough using real OpenStreetMap building heights extruded over terrarium DEM terrain. No Mapbox account, no tile bill, and a documented catalogue of the nine failure modes that compile cleanly and render an empty map.

CORX · Hindi broadsheet e-paper generator

Built a generator that turns an issue JSON plus images into a print-ready 355×560mm broadsheet PDF through headless Chrome, with per-zone word and character budgets so copy either fits the fixed type sizes or warns, plus an IST-aware daily scheduler that survives restarts without double-publishing.

Sahajyoga ERP · event and volunteer operations sahajyoga.vixitai.com

Built a 29-table event ERP — registration, payments, accommodation, volunteer assignment, check-ins — that deliberately imports nothing from the main platform, so it can be lifted to another server whole.

CricAI · live cricket scores and news cricai.site

Built a live score and news portal across 74 Express routes with Socket.IO scoring, web push over VAPID, and an admin CMS covering news, media, pages, matches, menus and SEO settings.

Infrastructure & Operations

Sole administrator of one Ubuntu VPS carrying 39 PM2 services, 17 PostgreSQL databases, 23 Nginx vhosts and 22 auto-renewed TLS certificates — every product and client site on this page.
Added 181 missing foreign-key indexes across production databases with no downtime, and set per-service connection caps after auditing oversubscription against the server’s limit.
Cut the public attack surface to ports 22, 80 and 443 in a hardening pass, moved secrets out of web-reachable paths, and pinned JWT algorithms after an authentication audit found an unpinned alg and an OTP bypass.
Found a backup that had been silently failing for five days while reporting success — offsite uploads were dead and the job still returned COMPLETE — and added the verification step that would have caught it on day one.
Fixed a client-settable X-Forwarded-For that defeated per-IP rate limiting across every site on the box, and run daily certificate-expiry monitoring with alerting after a renewal failure went unnoticed.
Document the reasoning behind infrastructure decisions in the repository, not just the configuration, so the next person inherits the argument rather than a file they are afraid to touch.

Competitions & Recognition

TII OrbitSight Challenge (US$50,000): submitted SPARSE as a reproducible container, with its results and its limits both stated in the artefact.
UNCTAD Supply Chain Innovation Challenge: authored “Chokepoint Sentinel,” a five-layer proposal for rerouting stranded cargo during trade-route disruption. A proposal and pitch site, not a deployed system.
Google Stitch Challenge: redesigned a full hospital-management interface, deployed live. Figma Config Makeathon: built “Keen Eye,” a design-skills arcade game, published to the Figma Community.
Hackathons: rapid-prototype entries for the Bubble hackathon (“PromptVault”) and a Spline competition (“Prompt Arena”). Brand systems: designed the complete VixitAI Labs identity — logo lockup, colour system and 30+ production-ready assets.

Technical Skills

AI & LLM: Chatbots and assistants on live sites; LLM APIs (Anthropic, OpenAI, Google, DeepSeek); prompt engineering (server-side system prompts, language overrides, structured output contracts); RAG and knowledge bases (pgvector, Voyage and Google embeddings, cross-encoder reranking, document ingestion); LLM-as-judge evaluation with a CI gate; transformers, torch and sentence-transformers behind FastAPI; AI-assisted development (Claude Code, daily)
Languages: TypeScript, JavaScript (Node.js and browser), Python, SQL, Solidity, HTML/CSS, Bash
Backend & APIs: Node.js, Express, Fastify, FastAPI, REST API design, BullMQ and Redis job pipelines, Socket.IO, webhooks, payment integration
Data: PostgreSQL schema design, forward-only migrations, query and index tuning, PgBouncer, PostGIS, pgvector, pg_trgm, Redis, multi-tenant isolation
Frontend & Mobile: React 19, Next.js 16, Tailwind CSS, Three.js and React Three Fiber, GSAP, React Native and Expo with over-the-air updates, Android APK and AAB builds, progressive web apps with service-worker strategies and VAPID push
DevOps & Infra: Ubuntu, Nginx, PM2 clustering, Docker and self-hosted registries, systemd, Cloudflare and R2, Certbot and TLS, deploy scripting, backups, monitoring and alerting
Security: Least-privilege database roles, JWT with token revocation, OTP and session auth, AES-256-GCM encryption at rest, audit logging, consent capture, PII masking, fail2ban, Nginx WAF rules, rate limiting
Testing & Practice: Vitest, Jest, e2e harnesses against isolated stacks, scripted release verification, PRDs and design documents written before code, Git, Figma, AI-assisted development (Claude Code, daily)

Education

Self-taught: four years of full-stack development, system design, databases, DevOps and application security — measured by what is running, not by coursework.
Comfortable being the only person who understands a system, and equally comfortable writing it down so that stops being true.
Higher Secondary Certificate (Class 12), CBSE, 2021.