Sethi Jyotir Aditya

Software engineer at a Big Four firm in Hyderabad, India, building production platforms, AI agent tools, and agent-driven Linux graphics work.

Interactive two-ink Risograph proof deck

Cobalt ink marks architecture and system scope Coral ink marks measured fixes and verified outcomes

Explore my work as a systems map →
Audit data platform operations, two-ink Risograph plate proof Two-plate Risograph schematic showing Cobalt architecture for roughly 20 repositories across Airflow, Databricks, AKS, Helm, and Vault overprinted with Coral fixes for 26 closed items, 5 of 6 Sev 1 and Sev 2 incidents, 66 restored users, and 41 down to 19 CVEs. ~20 repositories Airflow and Databricks Vault Unified AKS and Helm package (.NET 10 port) 41 to 19 CVEs 3 roles unmapped, 66 users 26 closed, 5/6 Sev 1 and 2

Production

Audit data platform operations

Read full spread

Scope: First-line engineering across roughly 20 repositories covering Airflow, Databricks, Python, .NET, Kubernetes/AKS, Helm, HashiCorp Vault, and Azure DevOps.

Problem: Multi-ticket Sev 1 lockouts, silent Airflow integrity failures, production Helm deployment collisions, and an unverified 41-CVE scan blocked routine platform releases.

Architecture: Unified Helm packaging, auth API recovery SOP, failure-reporting DAG hooks, and a .NET 10 container port from an audited IIS-to-AKS plan.

  • 26 work items Resolved or closed; 12 PRs merged across 4 repositories
  • 5 of 6 Sev 1 and Sev 2 Assigned incidents closed, including a 66-user access failure
  • 41 down to 19 CVEs OpenSSL scan across 8 binaries scoped to PyInstaller CVEs

Production

Audit data platform operations

26
work items resolved or closed
5 of 6
assigned Sev 1 and Sev 2 incidents closed
12 PRs
merged across 4 repositories

First-line production engineering across roughly 20 repositories spanning Apache Airflow, Databricks, Python, .NET, Kubernetes on AKS, Helm, HashiCorp Vault, and Azure DevOps.

Architecture and platform work

Operated across data ingestion pipelines, container infrastructure, and identity workers. Audited an IIS-to-AKS migration plan, documenting 7 concrete gaps before cutover, and ported the target application to a .NET 10 container build while the pilot sits on hold.

Consolidated split chart and container release definitions into a single unified Helm package so production upgrades apply configuration and image revisions atomically.

Measured fixes and incident outcomes

  • Traced an 8-ticket Sev 1 access outage affecting up to 66 users to a .NET worker mapping 3 roles onto 1 hidden permission as if 1:1; restored blocked users through the auth API and published the operations SOP.
  • Shipped an Airflow Parquet load failure-reporting fix and caught failed content-pack integrity checks that had been hiding behind green exit-0 Airflow tasks.
  • Resolved a production Helm chart-versus-image deployment collision by shipping one unified deployment package.
  • Audited a 41-CVE OpenSSL security scan across 8 binaries and scoped actual exposure down to 19 PyInstaller CVEs.

Production, live at 2 universities

University library platform

97%
of middleware authored (1,751 of 1,808 lines)
532 of 902
repository commits; 14 of 16 DevOps pipelines
>90% to <10%
live server CPU drop (under 30% during cron)

Campus library integration layer and deployment pipeline suite running in production across 2 universities, bridging Koha with institutional portals, identity, and finance systems.

Architecture and delivery

Authored 97% of the Node.js and Express middleware (1,751 of 1,808 lines; 532 of 902 repository commits) integrating Koha with campus portals, academic records, OIDC SSO, and idempotent finance billing.

Built 14 of 16 Azure DevOps pipelines along with roughly 30 theme overrides per campus, deploying cleanly to test, UAT, and production Ubuntu VMs.

Measured fixes and stability gains

  • Fixed 2 production-only permission-sync bugs across Koha's borrowers.flags and user_permissions tables.
  • Resolved a UAT OIDC SSO provider-name mismatch blocking federated logins prior to sign-off.
  • Blocked aggressive search crawlers in Apache, cutting live server CPU from over 90% to under 10% during normal load and keeping it under 30% during scheduled cron jobs.

UAT

Bank document Q&A chatbot

2.3 s vs. 6.1 s
cached versus fresh latency (0 LLM tokens on hits)
0.882 mean
RAGAS across 48 questions (0.91, 0.94, 0.78)
0 critical, 0 high
Checkmarx SAST across 28,677 lines; 184 tests

Main backend developer (186 of 207 commits) for a FastAPI payments-portal assistant, plus 22 Angular UI commits, engineered for strict banking security and air-gapped cluster deployment.

Three-tier retrieval and security architecture

Designed a three-tier query pipeline: preset matching at or above 0.70 similarity, a pgvector semantic cache at or above 0.92 cosine similarity (moved out of Redis 6.0.5 into Postgres and invalidated on 24-hour S3 ETag sync), and a role-filtered GPT-4o tool-calling loop capped at 5 steps and 10 calls over hybrid vector plus full-text Postgres search backed by 31 indexes across 9 tables.

Implemented raw ASGI AES-256-GCM body encryption using AWS KMS keys cached in Redis for 60 minutes, and packaged air-gapped Helm charts with baked tokenizer vocabularies.

Evaluation, masking, and security audit

  • Semantic cache hits respond in 2.3 s compared to 6.1 s for fresh runs while consuming 0 LLM tokens.
  • Integrated Microsoft Presidio masking tailored to Indian identifiers (Aadhaar, PAN, IFSC, UPI, bank and card numbers) with noisy PERSON and CVV recognizers disabled.
  • Measured RAGAS across 48 evaluation questions: 0.91 faithfulness, 0.94 recall, and 0.78 correctness (0.882 mean).
  • Passed bank Checkmarx SAST with 0 critical and 0 high findings across 28,677 lines, covered by 184 pytest and Hypothesis tests.

Internal tool and npm

AI agent tooling and MCP servers

288 runs
structured audit triage playbook executions
29 tools, 99 tests
agentic-playwright-mcp (58 E2E, 41 unit)
25 tools
waydroid-mcp on hidden fixed-DPI displays

Internal investigation tooling and two published open-source Model Context Protocol servers built to make daily browser, ticket, and Android automation safe for multi-agent workflows.

Playbooks, local pool, and MCP design

Authored a structured audit investigation playbook alongside 6 project skills that require explicit confirmation tables before any Azure DevOps write. Built a tokenless local Node.js Azure DevOps board (roughly 3,900 lines) backed by a 4-process MCP client pool holding 1,652 cached work items.

Published agentic-playwright-mcp, building on Patrick Menlove's ultimate-playwright-mcp and OpenClaw, for multi-agent work inside a shared signed-in Chrome browser using targetId routing, groupId tab groups, and automatic 3-second alert() dismissal.

Published waydroid-mcp exposing 25 tools for Android UI trees, input dispatch, screenshots, logcat, and Perfetto traces on hidden fixed-DPI displays.

Adoption and verification

  • Executed 288 structured triage runs using the confirmation-gated Azure DevOps workflow.
  • agentic-playwright-mcp exposes 29 tools verified by 99 automated tests (58 E2E and 41 unit) behind a coverage gate, and is used daily by Aditya and colleagues.
  • waydroid-mcp drives reproducible headless Android verification and Perfetto tracing without hijacking active desktop windows.

POC

Multi-tenant education chatbot platform

1,031 commits
of 1,464 total, plus 283 orchestrated agent commits
104 schemas
scoped to at most 5 tools per subagent
92 releases
scripted Angular and UMD widget builds; 2 MCP servers

Node.js and TypeScript multi-tenant conversational platform supporting isolated institutional configurations, hybrid document retrieval, structured analytics, and embeddable web widgets.

Multi-tenant and retrieval architecture

Built per-instance JWT verification and model key routing across 104 tool schemas kept tightly scoped to at most 5 tools per subagent. Supported 3 search modes: pgvecto.rs RAG, access-filtered Elasticsearch BM25+kNN reciprocal rank fusion, and MiniLM FAQ retrieval.

Added self-correcting read-only text-to-SQL generating D3.js charts, SSE reconnect replay for uninterrupted streaming, 92 scripted Angular and UMD widget releases, and 2 supporting MCP servers.

Security and reliability fixes

  • Migrated HR tools off direct SQL queries to a bearer-forwarding REST API so every call respects upstream user authorization.
  • Fixed broken video citations in a bilingual training demo by returning integer seconds directly in tool output instead of free-text timestamps.

Open source and Demo

Waydroid on NVIDIA Linux

3.6 to 0.5 cores
4K60 HEVC CPU via sandboxed FFmpeg/NVDEC
79 MB/s PCIe
down from 1.9 to 3.4 GB/s; jank 15 to 18% down to 1.7%
0/822 black frames
down from 608/740 via acquire-fence sync

AI-driven Linux graphics engineering and hardware validation on RTX 3060 and Hyprland, adding hardware Vulkan rendering, HDR10 video decoding, and desktop integration to Waydroid.

How it was built and stack architecture

Coding agents wrote most of the C and C++ code in separate threads for the Vulkan path, host decoder, composer, and per-app displays. In waydroid-venus, 53 of 61 commits carry a Cursor co-author trailer, and every change went through a second-agent review, typically Bugbot plus a second model, checking each finding against code or measurements before fixing or rejecting it.

Aditya's role was directing the work, reproducing failures, testing every build on his own RTX 3060 laptop, supplying observations such as black flashes and green frames, evaluating reviewer findings, and deciding what goes upstream and what stays on the fork.

The stack connects Android's Mesa Venus driver to virglrenderer Vulkan on NVIDIA's host driver with a sandboxed FFmpeg/NVDEC HDR10 host decoder and GPU-memory dma-buf buffers, crediting Shiro836/waydroid-nvidia for prior NVIDIA groundwork. Two fork items remain on hold awaiting upstream base pull requests: GPU-memory buffers and per-app displays (waydroid-app-displays, working on fork).

Measured GPU, decode, and upstream fixes

  • Sandboxed FFmpeg/NVDEC HDR10 host decoding dropped 4K60 HEVC CPU load from roughly 3.6 cores down to roughly 0.5 cores. All measurements come from one RTX 3060 laptop and one tester.
  • GPU-memory dma-buf buffers cut YouTube scroll PCIe traffic from 1.9 to 3.4 GB/s down to 79 MB/s and lowered frame jank from 15 to 18% down to 1.7%.
  • Traced green frames to a two-plane NV12 readback bug in NVIDIA's driver and filed a standalone C Vulkan reproducer on the NVIDIA developer forum.
  • Second-agent reviews from Bugbot and a GPT reviewer raised 13 findings across three Waydroid pull requests; 6 were fixed and 7 rejected with reasons after checking code or measurements.
  • Also merged a Koha batch-email plugin CSRF fix and an OpenCode documentation fix.
Upstream Waydroid pull request ledger (13 PRs: 3 merged, 10 open, plus review and fork notes)

Merged upstream (3 PRs)

Open upstream (10 PRs)

Review, fork queue, and other upstream fixes

  • Co-authored review accepted on open android_hardware_waydroid #91.
  • On hold awaiting base PRs: GPU-memory buffers and per-app displays (waydroid-app-displays, working on fork).
  • Additional merged fixes: Koha batch-email plugin CSRF fix and OpenCode documentation fix.

Verified credentials

Issuer-hosted records linked directly for verification.

Background and hiring

Hyderabad, India

Role
Software Engineer at a Big Four firm, based in Hyderabad, India.
Workflow
Most recent engineering runs through AI coding agents directed with repo playbooks, MCP tools, and second-agent reviews, with Aditya verifying claims, testing builds, and deciding what ships. The university library middleware was written by hand without coding agents.
Hiring
L1 technical interviewer for Python and AI engineering roles.
Education
B.Tech, Computer Science and Engineering, KIIT University, Bhubaneswar, CGPA 8.41.
Languages
English, Hindi, Telugu.

Grounded technical scope

Technologies exercised directly inside the six systems above.

Backend and APIs
Python (FastAPI, pytest, Hypothesis), Node.js, TypeScript, Express, .NET and .NET 10, REST, OIDC SSO, JWT, AES-256-GCM over raw ASGI, AWS KMS.
Data and retrieval
PostgreSQL (pgvector, pgvecto.rs, hybrid vector and full-text indexing), Elasticsearch (BM25+kNN RRF), Redis, Apache Airflow, Databricks, GPT-4o tool calling, RAGAS, Microsoft Presidio, MiniLM, read-only text-to-SQL.
Platform and ops
Kubernetes (AKS), Helm, HashiCorp Vault, Azure DevOps pipelines, Apache, Ubuntu VMs, air-gapped deployments, Checkmarx SAST.
Agent tools and web
Model Context Protocol (MCP), Playwright, Angular, D3.js, UMD widgets, SSE reconnect replay, Koha ILS.
Systems and graphics
Vulkan, Mesa Venus, virglrenderer, FFmpeg and NVDEC, dma-buf, HDR10 and HLG, Waydroid, Hyprland, Perfetto (C and C++ stack built with coding agents and validated on RTX 3060 hardware).