Portfolio Sethi Jyotir Aditya

Sethi Jyotir Aditya

Software engineer in Hyderabad. I build LLM applications, backend services, and AI agent tooling, and I support a live enterprise data platform.

Each line on this map is a discipline and each station is a project. Click any station for a concise summary, or jump straight to the work below.

Network map

  • Station on one line
  • Interchange
  • 1 openUpstream pull requests from that station, by status

Retrieval and LLM line

Getting a model to answer from the right documents, and checking whether it did.

Training video assistant

Demo, delivered to the client

A course assistant for a development bank's training on a farm loan process. It answers in English or Hindi and opens the training video at the moment that covers the answer.

The model kept citing minute 22 of a six-minute video. It was converting timecodes like 00:02:06 to seconds by hand and getting them wrong. I changed the tool output to integer seconds and hid the raw transcript timecodes, so the model copies a number instead of doing arithmetic.

  • Open Indian-language speech models covering 21 of the 22 scheduled languages.
  • 2 of 2 dev-server citation queries pointed at the right second. That is a smoke test, not an evaluation.

Quiz prototype

POC

Change here for Frontend line

A small app to test what a language model could do for a classroom. Upload a PDF and it writes a multiple-choice quiz, then scores you. Type a topic and it drafts a 45-minute lesson plan.

It ran on open-weight Llama 3 through a self-hosted Ollama server. The lesson I kept from it is that a model does not return clean JSON just because you asked.

Conversational analytics

POC

Change here for Frontend line

Ask "which quizzes did most students fail this term?" and get a chart back. The model writes SQL through tools that can only read. Either the database opens a read-only transaction, or only SELECT-type statements get through.

When a query names a column that does not exist, the tool looks up just the tables involved and their foreign keys, and the model fixes its own query.

  • Up to 3 correction retries per question.
  • SQL accuracy is unmeasured. No evaluation set exists for it.

Education chatbot platform

POC

Change here for Backend line Agent tooling line Frontend line

Retrieval in three forms. Document search on pgvecto.rs with each instance's data kept apart. Elasticsearch hybrid search over course pages, BM25 and kNN merged with reciprocal rank fusion and filtered by what the user may see. FAQ type-ahead from a self-hosted MiniLM embedding service.

Each subagent gets at most five tools. Past that, the model starts picking the wrong one.

  • 104 tool schemas in 10 skill packs.

Bank document Q&A chatbot

UAT

Change here for Backend line Platform and deploy line

A help chatbot inside a bank's fund-disbursement portal, for the government officials who create and approve payments. It answers from the official user manuals and cannot see or change transactions.

A question checks preset answers first, then a semantic cache of earlier answers, and only then reaches a GPT-4o tool-calling agent. The agent searches the manuals with pgvector and Postgres full-text search, filtered by the user's role. The cache matches at 0.92 cosine and clears whenever a manual changes.

Presidio masks Indian ID numbers before the model sees any text. Its default recognisers flagged ordinary portal terms as names and PINs, so I turned those two off and kept the structured IDs.

  • Cache hits used 0 LLM tokens and answered in 2.3 s against 6.1 s, on the test server.
  • RAGAS on 48 curated questions gave faithfulness 0.91 and answer correctness 0.78. The useful part was the three questions that scored near zero on retrieval.
  • GPT-4o judged its own answers. Next time I would use a different judge model.

Backend line

Services, schemas and integrations that other systems depend on.

Support-queue dashboard

Internal tool

Change here for Agent tooling line

A local Node.js board of my team's Azure DevOps work, with swimlanes by type and state and an activity page rebuilt from work item revisions.

The MCP batch call drops work item relations, so the board reads one item at a time. It runs a pool of 4 MCP server processes with 64 concurrent reads, retries 429 and 5xx responses with jittered backoff, and restarts a process when a sign-in page comes back instead of JSON.

  • About 3,900 lines.
  • A revision cache over 1,652 work items that fetches only revisions newer than its high mark.
  • Built for local team use without storing personal access tokens on disk.

Audit platform support

Production

Change here for Platform and deploy line Agent tooling line

A Sev 1 access failure on a Big Four firm's audit data platform. A portal returned an empty HTTP 500 when assigning users. I traced it to a 403 from the auth service, then to the platform's own .NET background worker.

Its sync loop mapped three roles onto one hidden role as if the mapping were one to one, so saving a role could switch the hidden role off. That is why the usual remove-and-re-add workaround worked only some of the time. The same symptom had produced eight earlier tickets without a diagnosis.

My first recommendation was a role grant. A control comparison against users who could still assign showed the only difference was one active flag, and that changed the fix.

  • Up to 66 users were in the same state. I restored access for the three blocked users through the auth API and published the procedure.
  • The code fix is committed on a branch and not yet merged.

Bank document Q&A chatbot

UAT

Change here for Retrieval and LLM line Platform and deploy line

The FastAPI service, as main developer. Chat streams over server-sent events, and Alembic migrations hold 31 indexes across 9 tables, 3 of them HNSW.

The government portal encrypts request bodies with AES-256-GCM, using keys from AWS KMS. Starlette's BaseHTTPMiddleware would not let me replace the request body, so I rewrote the middleware as raw ASGI that swaps receive.

The bank ran Checkmarx scans and I fixed what they reported. An AppSec finding on prompt input closed with a character allowlist.

  • 186 of 207 backend commits, and 184 test functions including Hypothesis property tests.
  • The bank's static analysis report showed 0 critical and 0 high findings across 28,677 lines.

Education chatbot platform

POC

Change here for Retrieval and LLM line Agent tooling line Frontend line

One Node.js and TypeScript backend serves many instances, each with its own agents, scoped tools, model keys and JWT key. A host app such as an LMS or an HR portal embeds the widget and says which instance and user role it is.

The widget calls the host's own domain and nginx forwards the request. The backend verifies the JWT against that instance's key, loads the skill, runs the tool loop and streams the reply over SSE. A stream survives a client disconnect and replays with 1, 2 and 4 second backoff.

The HR agent first spoke SQL to a database the chatbot host could not reach. I moved its five tools onto the HR service's REST API, forwarding the user's own bearer token, with a 15 second timeout and a distinct service-unavailable status.

  • 1,031 of 1,464 commits, plus 283 by coding agents I ran.
  • A wrong instance header does not fail. It runs the request against another instance's data, so it is the first thing I check.

University library platform

Production at two universities

Change here for Platform and deploy line Frontend line

Express middleware between Koha, the open-source library system, and each campus's student portal, academic records and finance system. It creates and updates library accounts, keeps staff permissions in line with portal roles, and sends fines to finance without billing anyone twice.

Koha stores permissions in two places, a bitmask on the user and rows in a permissions table, and both have to match. Real users found two bugs that UAT missed. Users with a null flags value were never compared, and the sync overwrote librarians an admin had restricted by hand.

  • About 1.8k lines, 97% of the middleware source.
  • Today I would replace the string-built SQL in the fines updates with parameters.

Platform and deploy line

Getting code onto servers and clusters, and working out why it broke there.

Agency site and CMS

Website in production, CMS in UAT

Change here for Frontend line

A static Next.js site for a marketing consultant's agency. Its contact form is a Hono function on Cloudflare Pages that writes to Neon Postgres behind a Turnstile check.

For his own site, a block CMS. The admin runs on one Cloudflare Worker behind Cloudflare Access and edits blocks in Postgres. A render step writes them into the page before the build, so the site stays static and still builds without the database.

The Neon HTTP driver has no interactive transactions. Every write is one SQL statement built from data-modifying CTEs, and a stale version number returns 409, so two open tabs cannot overwrite each other.

  • 21 production deploys of the agency site through GitHub Actions.
  • The CMS round-trip check rebuilds 15 homepage sections and 39 posts byte for byte.

University library platform

Production at two universities

Change here for Backend line Frontend line

Azure DevOps pipelines for the middleware and themes. A build script bundles the middleware with esbuild and commits it to an environment branch, and each branch triggers a deployment job on its test, UAT or production Ubuntu VM.

A production server ran hot. I traced the CPU to search engine crawlers working through catalogue search and blocked them from search in the Apache config. CPU fell from over 90% to under 10%, and stayed under 30% during cron jobs.

A UAT sign-in failure came from configuration, not code. The identity provider's redirect URIs used production provider names on a UAT server.

  • 14 of the 16 pipeline definitions.
  • VM sizing and DNS, SSL and firewall requirements for two universities.

Waydroid on NVIDIA

Demo, public repo

Change here for Agent tooling line

It started as a way to play Apple Music lossless on Linux, through Apple's Android app in Waydroid. On my NVIDIA laptop, Waydroid drew Android on the CPU.

Android's Mesa Venus driver now sends Vulkan calls to virglrenderer on the host, which runs them on NVIDIA's driver. Around that sit a host video decoder, HDR output, composer fixes, and per-app displays. The per-app displays work on my fork and are on hold until the related upstream changes land, so they can be proposed on top of them without conflicts. Ideas ported from Shiro836's waydroid-nvidia, a community NVIDIA build, keep sampled buffers in GPU memory.

  • 4K60 HEVC for about half a CPU core, against about 3.6 in software.
  • Janky frames on a YouTube scroll from 15 to 18% down to 1.7%, close to Shiro836's build on the same test.
  • One machine, an RTX 3060 laptop GPU, and one tester.
Upstream status 3 merged, 10 open pull requests

Each fix I send to Waydroid is its own pull request. The list also holds fork work, the NVIDIA report and a review.

Audit platform support

Production

Change here for Backend line Agent tooling line

An Airflow fix so that Parquet load failures report Error to the orchestrator instead of leaving requests stuck. I took it through the UAT, Stage and Prod release runbooks.

At Prod, another runbook targeted the same service with a newer Helm chart carrying an older container image, so whichever deployed last would undo the other. Only inspecting the image commit showed which code each package carried. One combined Helm package aligned both runbooks onto the same chart and fixed image.

A content pack once failed its integrity tests while the pod exited 0 and Airflow showed the task green. Data-correctness tickets now start with the task log.

  • 26 support items resolved or closed.
  • 12 merged pull requests across four repos.

Bank document Q&A chatbot

UAT

Change here for Retrieval and LLM line Backend line

Dev, UAT and prod Helm charts and the container build for a Kubernetes cluster with no internet egress. tiktoken downloads its vocabulary on first use, so the image bakes it in, with a character-count fallback.

The design put the semantic cache in Redis, and the bank's Redis 6.0.5 could not hold it. I moved the cache into the Postgres instance that already ran pgvector, which removed a dependency and put the cache next to the ingestion job that clears it.

For an internal demo I stripped out bank-specific fixtures and host names and packaged a Docker Compose stack with an Azure Pipelines deploy.

  • Baked tokenizer vocabularies and parameterized mirrors for air-gapped cluster builds.

IIS to AKS port

Pilot, on hold

A .NET web app ran on Kubernetes in lower environments and on Windows IIS in UAT, Stage and Prod. I checked the migration plan against the live pipelines, Vault, clusters and Datadog.

The plan had seven gaps. The image promotion step tagged the wrong registry path, the secret list was missing four sign-in keys, and the QA worker logged 615 Event Hub timeouts in 24 hours. Then I ported the container build to .NET 10 on a release branch.

  • 24 commits on the port.
  • NuGet advisories went from 6 to 0 on the branch, which is not merged.

Agent tooling line

Tools that let coding agents read, browse and act, with a person approving every write.

Research review tools

Internal tool

Change here for Frontend line

For a friend's research on how LLMs make judgement calls, an MCP server lets a coding agent generate benchmark scenarios. It writes each one to disk only if it passes a strict schema, so a malformed scenario never reaches the dataset.

Figma agent bridge

Demo

Change here for Frontend line

A Node MCP server that talks over a WebSocket to a Figma development plugin. The agent can send messages, run plugin JavaScript against the document, and export frames as PNG to look at its own layout.

  • A school group's 83-KPI spec became eight branded dashboard screens.

Education chatbot platform

POC

Change here for Retrieval and LLM line Backend line Frontend line

Two MCP servers expose the platform's tools to coding agents. One is a local stdio server. The other is a stateless Streamable HTTP server inside the backend, with a skills tool and an audit table of every invocation.

I ran coding agents on the platform through a tracker of 161 work streams. An agent had to claim a stream and its allowed paths before editing, and record how it verified the work.

  • 283 agent-authored commits, all pushed from my account.
  • Most tracker rows record a build check, not a test run. A full suite run after one deploy still had 18 failures out of 228.

Chatbot mobile client

POC

Change here for Frontend line

An agent-built prototype client for the education chatbot platform. A shared Kotlin module holds the domain models, MVI contracts and ViewModels, a Ktor REST client, an SSE streaming client, repositories and Koin dependency injection.

Built an Android debug APK inside a Docker dev container alongside a Wasm web preview for rapid UI verification.

  • About 8,000 lines across 121 Kotlin files (Android APK & Wasm preview).

Waydroid on NVIDIA

Demo, public repo

Change here for Platform and deploy line

Coding agents wrote most of this code, in separate threads for the Vulkan path, the decoder, the composer and per-app displays. Every change had at least one review by a second agent, usually Bugbot plus a second model, and the agents checked each finding against the code or a measurement before fixing or rejecting it.

My part was the goals, testing every build on my own laptop, reporting what I saw, such as black flashes and green frames, and deciding what goes upstream.

  • 53 of the 61 commits in waydroid-venus carry a Cursor co-author trailer.
  • Bugbot and a GPT reviewer raised 13 findings on three of the Waydroid pull requests. 6 were fixed and the rest rejected with reasons.

Audit platform support

Production

Change here for Backend line Platform and deploy line

Agents read the work item history, logs and code, and draft my analyses. I choose what to investigate, check each claim against the evidence, and decide what goes to a ticket, a reviewer or production.

They reach Azure DevOps through its MCP server, and the sign-in-protected tools such as Datadog, Airflow and the wiki through agentic-playwright-mcp in a signed-in browser.

  • 35 of my 60 commits on the platform carry an agent co-author trailer.

Support-queue dashboard

Internal tool

Change here for Backend line

The board holds no access token. It starts the same Azure DevOps MCP server my agents use and talks to it as an MCP client, so sign-in is whatever that server already does and the app keeps no secret.

The cost is one call per work item, which is why it needs a process pool and a cache to stay fast.

Investigation playbook

Internal tool

A written playbook my agents follow on the audit platform. It says how deep to go, where each kind of log lives in each environment, and how to prove a fix is live.

Most rules record an incident. A log query with zero results does not prove a code path did not run. Run a control query before trusting an onset date. Search for earlier cases of the exact error before calling anything a one-off.

Six project skills cover release runbooks, production hotfix scripts, and turning a production issue into a team bug. Every skill that writes to Azure DevOps displays a confirmation table and blocks until approved—a safeguard introduced after an early unreviewed draft triggered unwanted team notifications.

  • A 57 KB playbook.
  • 288 agent runs across 34 threads on the audit platform.

agentic-playwright-mcp

Internal tool, published on npm

Stock Playwright MCP gives one agent one implicit tab, so two agents on one browser drive each other's pages. This server lets several agents share one signed-in Chrome. Every tool takes a tab id, and each agent lists and closes only the tabs in its own group. A person signs in once and every agent carries on in that session.

One unanswered alert() froze every agent on the shared Chrome, so the server now dismisses dialogs after about three seconds.

It builds on Patrick Menlove's ultimate-playwright-mcp and code from OpenClaw. I use it daily, and colleagues run it for their own agents.

  • 29 tools, 58 end-to-end tests and 41 unit tests.
  • A coverage spec fails the suite when a tool has no test.

waydroid-mcp

Internal tool, published on npm

An MCP server that lets a coding agent run and debug Android apps in Waydroid. It installs and opens apps, takes screenshots, reads the UI tree, taps and types, follows logs, and collects crashes, stack dumps and Perfetto traces.

With waydroid-app-displays, an app runs on a hidden display of a fixed size and density, so an agent can check a 360 dp phone layout without opening a window on my desktop.

  • 25 tools covering app lifecycle, UI hierarchy, input, logs, and Perfetto traces.

jev-browser-sidekick-mcp

Demo

A side project that lets a coding agent hand off simple browser chores in plain words. For each step it lists the controls that could do it and asks Jev, a hosted decision model from TypeSafe, to pick one or say none fits. It drives the same Chrome as agentic-playwright-mcp, so the calling agent can pick up any tab where a step stopped.

My own test report found that when the requested button was missing, the server clicked something else at confidence 1.0. The fix was a standing "none" option in the choices, not a different model.

  • 8 versions on npm and 56 test cases.
  • 565 of 589 decisions succeeded in development traces, at about $0.00005 each.

Frontend line

The parts people use directly: widgets, themes, dashboards and review apps.

Research review tools

Internal tool

Change here for Agent tooling line

Two small React and Ant Design review apps, so people can label the data without touching JSON. A small Express API writes each reviewer's answers to files, and both apps run on my own server behind Apache and Cloudflare.

One app hides which model wrote each conversation and checks reviewers with a calibration step first.

  • Two reviewers scored 100 items each.

Figma agent bridge

Demo

Change here for Agent tooling line

Eight dashboard screens in HTML under a client's brand guide, covering prospects, students, employees, finance and leadership. The bridge then carried them into Figma.

  • 78 screenshots from the fix-and-check rounds.

Quiz prototype

POC

Change here for Retrieval and LLM line

Three classroom tools in one Angular app. A quiz from a PDF with scoring, a lesson-plan draft from a topic, and a mood check-in that answers more gently when the rating is low.

Conversational analytics

POC

Change here for Retrieval and LLM line

The chart side. The model reshapes a large result into at most 15 points, and chart services in the chat widget draw them with D3.

  • About 1,050 lines of widget chart code.

Education chatbot platform

POC

Change here for Retrieval and LLM line Backend line Agent tooling line

An embeddable Angular chat widget. A release script publishes it as Angular 14, 18 and 19 library builds and as a UMD bundle, installed in dev LMS sites and an HR portal branch. Documented the HR portal's vendored Angular version pinning so teams can track drift cleanly.

  • 92 releases.

Chatbot mobile client

POC

Change here for Agent tooling line

A mobile client for the education chatbot platform. One Compose Multiplatform UI targets Android, iOS, Kotlin/JS and Wasm. You set up an instance by scanning its QR code with CameraX and ML Kit, then sign in through SSO deep links. Replies render as Markdown that stays valid while it streams, including tables and the backend's frontend-tool output.

Testing ran on a physical phone or in a Wasm preview inside a phone-sized frame. The preview reaches the backend through a same-origin proxy and caches with a service worker.

  • Android debug APK and Wasm preview verified; prototype scope only.

University library platform

Production at two universities

Change here for Backend line Platform and deploy line

Branded catalogue and staff interfaces for each institution. Sign-in hands off to the university's SSO, new users get a password-set prompt, and librarians get dashboard charts. I shared the theme work roughly evenly with a teammate.

  • About 30 template overrides per site.

Agency site and CMS

Website in production, CMS in UAT

Change here for Platform and deploy line

GSAP scroll effects for the agency site, including pinned horizontal service cards with 3D reveals and a persistent particle canvas. A performance pass stops canvases once they leave the page and pauses animation in hidden tabs.

The CMS admin has a block editor with live preview, a rich-text blog editor, drag to reorder, and release history with rollback.

  • 6 Playwright specs on the site.
  • A pixel comparison against the live site showed 0.00% difference at 1440 and 390 px wide, apart from the captcha frame.

About

I'm a software engineer in Hyderabad, at a Big Four firm. I build LLM applications and backend services, and I work production problems on a live data platform.

Most of my engineering runs through AI coding agents. I write the playbook an agent follows, connect it to Azure DevOps and a signed-in browser through MCP servers, and review what comes back before it reaches a ticket, a reviewer or production. The library middleware on the Backend line I wrote without agents.

Education
B.Tech, Computer Science and Engineering, KIIT University. CGPA 8.41.
Languages
English, Hindi and Telugu.

Certifications

Open source

agentic-playwright-mcp is a Playwright MCP server for several AI agents sharing one signed-in Chrome. It builds on Patrick Menlove's ultimate-playwright-mcp, and the README credits him and the OpenClaw code it uses.

waydroid-venus runs Android's Vulkan calls in Waydroid on NVIDIA's host driver, and adds a host video decoder. The Waydroid fixes I send upstream go one fix per pull request, and I reported a driver bug to NVIDIA with a reproducer. The Waydroid stop lists each one with its status.

waydroid-mcp is an MCP server that lets a coding agent run and debug Android apps in Waydroid, with 25 tools for installing and opening apps, screenshots, the UI tree, input, logs, crashes and Perfetto traces. With waydroid-app-displays, an app can run on a hidden display of a fixed size and density, so an agent can check a 360 dp phone layout without opening a window on the desktop.

Smaller contributions: a CSRF token fix, merged, that makes the National Library of Finland's Koha batch-email plugin work on newer Koha releases, and a merged docs fix in OpenCode.

Contact

The quickest way to reach me is a message on LinkedIn.