Ayush Patravali

Journey

What I owned at each place, newest first, with one real number or diagram per project.

1. Apr 2026 to now

   Ramco Systems

   AI Engineer

   Bengaluru

   Enterprise HR and payroll software. I design and ship internal platforms end to end, using AI coding agents under a review loop I run.

   Aug 2026 to Oct 2026

   Self-service payroll data integration platform

   Payroll customers needed their data in banks, finance tools and benefits systems, and every integration was custom work. I designed and built a platform that lets them set it up themselves.

   1. API gatewayScoped credential per client, mutual TLS
   2. IdentitySingle sign-on, invites, dynamic roles
   3. Integration backendWizard, scheduler, job queue, delivery
   4. EmailSend-only notifications
   5. Web appThe six-step connector wizard

   My roleLead developer, designed and built it end to end

   • Five services from scratch, an API gateway, a reusable identity service, the integration backend, an email service and the React app.
   • Every client application gets its own scoped credential behind the gateway, and callers prove their workload identity over mutual TLS, so a leaked secret can't be replayed.
   • Scheduled, encrypted file delivery built on PostgreSQL row locks instead of adding a message broker, with calendar rules and a preview of the next runs.

   Running in a UAT environment after about two months, with around 200 design decisions recorded along the way.

   • Go
   • React
   • TypeScript
   • PostgreSQL
   • Kubernetes
   • OIDC

   Read the case study

   Apr 2026 to Aug 2026

   Automated configuration audits for global payroll

   Payroll consultants checked each customer's configuration by hand, one SQL query and one spreadsheet at a time. I built a platform that turns their checks into reusable rules and runs them on a schedule across every customer and country.

   Before15 to 18 min

   Afterabout 2 min

   My roleSole developer, with input from the payroll experts

   • Rules written once run on a schedule across every customer, country and operating unit, read-only, with failures landing in a review queue with the evidence attached.
   • Made full-data pulls complete and about 8x faster with parallel, rate-capped fetching.
   • An AI assistant drafts rules from plain English against the real column list in a strict schema; a person always approves the draft.

   Large audits about 8x faster, and every finding traceable to the rule, the data and the reviewer.

   • Python
   • FastAPI
   • React
   • TypeScript
   • SAML SSO

   Read the case study

   Jun 2026 to now

   AI payroll operations automation, deployed on AWS

   Payroll teams repeat the same 15 to 20 manual checks every month for every client. The platform has AI write each check once as code, a person approves it, and it replays every month. I led its AWS deployment.

   1. AI writes the checkFrom the operator's plain-English description, as tested code
   2. A person approves itNothing runs without review
   3. It replays every monthNo AI cost until the input files change shape

   My roleCore engineer on a five-person team; led the AWS deployment

   • Led the AWS deployment and release pipeline, containerised services, CI pushing images to a registry, a pull-based deploy, object storage and managed data stores.
   • Enterprise single sign-on and hardened sign-in, including multi-factor login for admin accounts.
   • Rebuilt the operator console on a shared design system, fixed queue reliability, and shipped an AI email agent that only sends after a person reviews it.

   Deployed on AWS, with every AI-written automation approved by a person before it runs.

   • AWS
   • Docker
   • Node.js
   • React
   • Redis
   • MongoDB

   Read the case study

2. May 2026 to now

   Side project

   Independent software

   Karnataka

   Kannada loan paperwork, automated for a rural co-operative bank

   Field officers filled loan applications into old Excel templates in a legacy Kannada font, then printed them page by page. I built software that turns a validated form into the bank's full official loan packet.

   6

   loan schemes

   17 to 23

   pages per packet

   85

   render test cases

   My roleSide project, solo

   • A validated web form with the bank's own calculations for crop income, land valuation and instalments, in Unicode Kannada only, so garbled legacy text can never reach a printout.
   • Prints the bank's official 17 to 23 page loan packet as a PDF in a second or two, across six loan schemes.
   • An 85-case render test matrix covers every scheme, and the on-premise install has signed offline licences, daily backups and automatic rollback.

   Pilot-ready; not yet installed at a bank.

   • React
   • FastAPI
   • SQLite
   • WeasyPrint

   Read the case study

3. May 2025 to Mar 2026

   National Institute of Advanced Studies

   Research intern

   IISc campus, Bengaluru

   Neurochaos learning, extended to graph data

   Neurochaos learning learns from small samples on modest compute, but had only been tested on flat and time-series data. We extended it to graph data for the first time.

   1. Cora0.844
   2. PubMed0.76
   3. CiteSeer0.732
   4. Wisconsin0.57
   5. Cornell0.43
   6. Squirrel0.34
   7. Actor0.265

   My roleResearch intern, second author

   Preprint on arXiv (Feb 2026). Macro-F1 0.844 on Cora, falling to 0.265 on Actor as neighbours' labels diverge.

   • Python
   • PyTorch Geometric
   • NumPy
   • scikit-learn

   Read the case studyarXiv:2602.16204Code

4. Sep to Dec 2025

   Suketa Technology Solutions

   Intern

   Bengaluru

   Hybrid retrieval for large, mixed-format document sets

   Finding the right passage in a large corpus that mixes medical papers, threat reports and law, across PDF, Word, CSV and JSON. We built a retrieval pipeline that ingests it all and returns the best five passages.

   1. Every chunkPDF, Word, CSV, JSON; type-aware chunking
   2. 200 candidatesBM25 + dense search, fused with RRF
   3. Top 5Cross-encoder rerank

   My roleIntern; two-person build, I owned chunking and the pipeline documentation

   Raw files in, five ranked passages out per query, built with one other engineer.

   • Python
   • Milvus
   • BM25
   • Sentence Transformers
   • Docker

   Read the case study

5. Aug to Dec 2024

   Aytasense Technologies

   R&D intern

   Bengaluru

   Real-time iris-to-pupil ratio detection

   The iris-to-pupil ratio is a possible non-invasive marker for neurological conditions. We built a pipeline that measures it from a camera frame in real time.

   0.987

   mAP@0.5

   0.989

   precision

   0.949

   recall

   My roleR&D intern, team of four

   Precision 0.989 and recall 0.949 on validation; paper submitted to Biomedical Materials and Devices.

   • Python
   • YOLOv5
   • OpenCV
   • Streamlit

   Read the case studyCodeApp

6. 2022 to 2026

   RVITM

   BE, Computer Science and Engineering, with my final-year project

   RV Institute of Technology and Management, Bengaluru

   Sanjeevani, multi-agent RAG for Indian medicinal plants

   Questions about medicinal plants span pharmacology, where a plant grows and whether it is endangered. One retriever can't answer all three, so a router sends each question to a specialist agent with its own index.

   ResearchGISIUCN

   Research1500

   GIS2130

   IUCN1014

   My roleFinal-year project, team of four; I built the agent system

   The router sends 42 of 45 test questions to the right agent, and retrieval answers in a median 0.12 s.

   • LangGraph
   • Weaviate
   • Llama 3.3
   • Whisper
   • Streamlit

   Read the case studyPaper (IEEE)Code