Krishan – LLM, LangChain, AI agent development

Decision intelligence software for manufacturing SMEs. It connects the systems a company already runs, including ERP, CRM, email, and contracts, into a single retrieval layer, then runs agents over that layer to surface margin erosion, supplier risk, and compliance gaps as specific alerts instead of dashboards. Access is read-only, and deployment takes about two weeks, so a customer needs no data migration.

• Designed the retrieval architecture, combining time-aware ranking, parent-document expansion, and automatic token budgeting that decides how much context each query needs, keeping answers accurate as the corpus grew and holding token cost per query flat.
• Built an autonomous agent layer on top of the retrieval engine that plans and executes multi-step investigations across company data, with per-agent security policies bounding what each agent can read and change. Hence, agents run on sensitive business data without a blanket access grant.
• Built ingestion connectors for ERP, CRM, email, contracts, Notion, GitHub, Dropbox, and Google Drive, so a customer can connect read-only to the tools they already run and get answers in about two weeks with no data migration.
• Grew the platform from zero to 3 paying companies and around 12 production users, running discovery, build, and sales directly with customers.
• Ran the full stack on a high-availability GCP Kubernetes cluster, supporting production customers with no operations staff.

A portfolio site where the whole interface is an AI agent. Instead of browsing pages, a visitor talks to an agent who answers in first person from a retrieval layer built over a career history, and the agent drives the interface. In contrast, it answers by switching the theme to match the topic, changing language, opening the calendar, and triggering the CV download. Every answer is grounded in retrieved content, so the agent refuses any question for which it has no source.

• Built a tool called an agent that controls the front end while it answers, switching theme, language, sound, and calendar state through typed actions, so the interface reacts to the conversation instead of to navigation.
• Grounded every response in a retrieval layer over the career corpus and blocked generation outside it, so the agent states no fact about a real person that it cannot source.
• Shipped bilingual handling that detects the language of each incoming message and switches the entire interface to match, with no user setting and no page reload.
• Generated three contextual follow-up suggestions after every response, so a visitor keeps moving through the conversation instead of stopping at one question.

A B2B sales CRM for small and mid-size sales teams. It brings pipeline management, calling, email, and SMS into one workspace, so a sales rep can run the entire cycle without leaving the tool. The product spans a large single-page web application and a mobile app that gives reps the same pipeline, inbox, and calling features in the field.

• Drove Mobile app delivery of a WebView/React Native hybrid under two product constraints, no rebuild from scratch and no changes to the web codebase, cutting user-reported open bugs on the Play Store by over 80% within 2 months.
• Shipped the company's first mobile app, owning the implementation cycle and the mobile team end-to-end, and stabilizing the WebView-to-native bridge that caused most user-reported failures.
• Rebuilt the mobile release process around the hybrid architecture, so web changes reached mobile users without a native release and without the regressions that drove the original bug count.
• Converted the codebase from JavaScript to TypeScript, removing a whole class of runtime type errors from the release cycle.
• Migrated hundreds of components from Backbone.js to React in a live production application.
• Built the company's first agentic UI layer, an LLM agent that executes actions inside the user's interface. Hence, a rep completes a multi-step CRM task by asking for it instead of navigating the UI.
• Shipped the company's first AI workflow generator: an agent that turns a plain-language request such as "follow up with a lead after three days of silence" into a complete executable automation, letting non-technical sales users build their own automations without engineering support.

A three-sided marketplace for vacation rental tourism. Rental hosts hand their guests a personal link at check-in; guests unlock discounts at restaurants, shops, and experiences near the property; and local vendors reach tourists who are already in the area and verified as staying there. Hosts pay nothing, and the platform takes a 5% fee on each vendor transaction.

• Built and shipped a three-sided marketplace to production in five months as one of two founders, covering host onboarding, the vendor offer console, and the guest-facing mobile experience.
• Designed a geo-targeted flash offer system where a vendor with idle capacity pushes a time-limited discount to every app user inside a radius, and the first guest to accept gets the highest price, converting unsold restaurant covers into same-day revenue.
• Closed and integrated a partnership with the region's largest tourism body, which supplied the initial supply side and brought the platform to 40 properties and 15 vendors at launch.
• Ran the full stack on Kubernetes with a two-person team, serving three distinct user types with no operations staff.
• Led 2 developers alongside the build to compress the delivery timeline, setting the architecture and reviewing the code they shipped.

A tablet-based ordering and inventory platform for restaurants. Diners order from a digital menu at the table, the order prints in the kitchen through a thermal printer, and stock levels update in real time as items sell. Built for small independent restaurants that wanted table service without a waiter taking every order.

• Built the platform end-to-end as technical co-founder, covering the diner ordering interface, the kitchen printing pipeline, and the real-time inventory engine, taking it from nothing to a working pilot in 4 months.
• Integrated a thermal printer SDK so that a table order reached the kitchen printer with no staff involvement, removing the waiter handoff that the whole product depended on.
• Designed the real-time data layer on Firebase so menu changes, orders, and stock counts stayed in sync across diner tablets, the kitchen, and the manager dashboard without polling.
• Led 2 developers, setting the architecture, reviewing code, and owning delivery.

A time tracking product built around a physical eight-sided device. The user flips the device onto a face, and the desktop and mobile apps start tracking that activity, then turn the tracked time into reports and timesheets. The apps run on Windows, macOS, iOS, and Android and pair with the device over Bluetooth.

• Owned architecture and feature work across the Electron desktop app and the React Native mobile app inside an eight-person engineering team, holding both clients at feature parity while the product shipped on four operating systems (iOS, Android, MacOS, Windows).
• Rewrote the Bluetooth pairing and reconnection layer between the apps and the physical tracker, which was dropping connections mid-session, draining device battery, and crashing the client, and removed all three as recurring user complaints.
• Led multiple refactors of the shared client codebase with no feature freeze, so releases continued while the architecture changed underneath.

A consumer fitness tracking app for running and training on Android and iOS, with a social feed where users follow friends, share activities, and receive coaching content. The Android client was fully native, and the team was introducing React Native into it screen by screen. The app served roughly 15 million monthly active users.

• Led the integration of a React Native news feed into a large native Android app, writing the custom native modules, managing the bridge, and coordinating lifecycle between native activities and React Native views. Hence, the feed matched fully native screens in performance.
• Reverse-engineered a decompiled React Native Java APK to trace a defect that had been open on the React Native repository for over five years and hit a specific subset of users, then shipped a client-side workaround that handled it without waiting for an upstream fix.
• Worked across JavaScript, TypeScript, Java, and Kotlin daily inside a five-person cross-functional team, shipping into an app with roughly 15 million monthly active users.

A blockchain technology company with two business lines. The first was a cryptocurrency mining operation of 30 rigs and 180 GPUs that switched between coins as profitability moved. The second was B2B blockchain consulting for manufacturing companies seeking supply chain traceability, which grew to more than 10 significant clients. Internal products included a farm-to-fork traceability platform that wrote sensor readings from wine production onto Ethereum.

• Built and ran a 180-GPU mining operation across 30 rigs, architected to switch coins as profitability moved and to run unattended, keeping it profitable for roughly 18 months with minimal intervention.
• Pivoted the company from mining to B2B blockchain consulting when computing cost overtook mining margin, and grew that line past 10 significant manufacturing clients and €800k in revenue.
• Hired and led an engineering team of 12, five mid-level and seven junior developers, owning hiring, code review, and the standards for scalability and code quality.
• Delivered a farm-to-fork traceability platform on Ethereum that wrote production sensor data on the chain, taking it to 3 beta wineries.

A software agency delivering custom web and mobile applications for external clients. Work covered business web applications with PHP backends and JavaScript front ends, and cross-platform mobile apps for Android and iOS, each scoped and delivered directly to the client who commissioned it.

• Delivered client web applications end-to-end on a PHP and JavaScript stack, taking each one from database schema through to the delivered interface with no separate backend or frontend owner.
• Built cross-platform mobile applications in Ionic and native Android, gathering requirements directly from the client and shipping within fixed agency deadlines.
• Shipped more than 4 client projects in seven months across four languages, moving between PHP, JavaScript, Java, and Android within the same week as project deadlines demanded.

A two-sided fitness platform. Gyms managed memberships and monitored live occupancy in the building. Personal trainers built workout plans, tracked client health data, and monitored progress via a dashboard that synced with each client's phone. Members used it for free, and gyms paid for a subscription. It also carried gym announcement channels and a live count of how many people were training at that moment.

• Built the entire platform solo while at university, covering the native Android app for members, the React dashboard for gyms and trainers, and the PHP backend serving both.
• Built a live gym occupancy view and gym announcement channels on top of the membership data, giving independent gyms a social layer that the large fitness apps did not offer.
• Shipped a two-sided subscription model, free for members and paid for gyms, and took the product to public beta.
• Ended the project when an international fitness app company offered a full-time engineering role on the condition that the product left the market, which started a two-year role there.