Software Engineer II · JPMorgan Chase · Bangalore

Hi, I’m Ayan Sadhukhan. I build search and agent infrastructure at JPMorgan Chase.

I owned the RAG search service behind chase.com’s search bar, end to end, and now build features for the agentic operating memory behind Chase’s multi-agent AI assistant. Python, Java and AWS. Every claim on this page comes from my versioned work log, numbers included.

See the work Daybook, my side project
vault.db › achievements › #15 Oct 2026

Agentic Operating Memory

Solely built the service's gRPC interface alongside its REST API in one codebase, cutting p90 response latency from 30 ms over REST to 12 ms over gRPC

p90 response latency, REST vs gRPC
30 → 12 ms −60% computed

skills gRPC · Java · Protocol Buffers · REST APIs

1 / 4status: active · impact 3/5
18Mchase.com customers served by the RAG search service I owned
~4Mrequests a day to FAQ Search, which replaced a $200K/yr vendor product
30→12 msp90 latency after I built the service's gRPC interface
25→5 minindexing latency with concurrent execution in AWS Lambda

01 / Work

Nearly four years at JPMorgan Chase, from intern to SE II.

JPMorgan Chase & Co.

Bangalore, India · Feb 2023 – now

  1. Software Engineer IIFeb 2026 – now
  2. Software EngineerJun 2023 – Jan 2026
  3. Software Engineer InternFeb – May 2023

Jan 2026 – now

Software Engineer II

Agentic Operating Memory

Agentic operating memory for the Chase AI agent: session, conversation and profile-context APIs

30→12 msp90 latency, REST → gRPC
~1,000employee beta testers in the pilot
5environments, dev through prod
  • Implemented features for the agentic operating memory behind Chase's multi-agent AI assistant: a Java Spring Boot service serving PII-masked customer context, piloted with ~1,000 employee beta testers
  • Implemented features on its memory layers: Redis for short-term session memory and Cassandra for storage
  • Solely built the service's gRPC interface alongside its REST API in one codebase, cutting p90 response latency from 30 ms over REST to 12 ms over gRPC
  • Implemented features contract-first against the team's published API contract, writing most of the code with AI assistants (Claude, GitHub Copilot)
Show 3 moreShow less
  • Shipped the service through dev, test, UAT, pre-prod and prod, backed by unit, Karate, integration, acceptance and end-to-end test suites
  • Ran the service on Amazon EKS behind Route 53 and Envoy routing, with health endpoints that gate traffic on downstream status and Splunk, Dynatrace and Datadog monitoring
  • Built a small internal UI to demo the agentic operating memory's session, conversation and context APIs
  • AI-Assisted Development
  • Context Engineering
  • Java
  • REST APIs
  • Spring Boot
  • AI Agents
  • Amazon EKS
  • Amazon Route 53

Aug – Dec 2025

Software Engineer

Galileo Guardrails

Onboarding the Galileo AI guardrails platform into the firm with the Galileo team

  • Partnered with the Galileo team to onboard its AI guardrails platform into the firm, bringing it onto the team's Amazon EKS cluster
  • Amazon EKS
  • AWS

Oct 2024 – Dec 2025

Software Engineer

RAG Content Search Service

RAG-based generative AI search service behind the chase.com search bar (18M customers), built in FastAPI behind a search orchestrator, answering with citations, the matching section of a page or PDF, or a generated answer. Kickoff to production (Jul 2025), supported until Dec 2025

18Mchase.com customers
1.5–2Mrequests a day
300 msp90 response latency
25→5 minindexing latency
  • Owned a RAG search service for 18M chase.com customers (1.5–2M requests/day) end to end while covering for the team lead: wrote its API contract and low-level design, built it in FastAPI; 300 ms p90
  • Designed the service's first ETL pipelines (AWS Lambda, Step Functions, OpenSearch), replacing manual indexing with cron-scheduled runs that scrape and embed web and PDF content; halved, then eliminated, downtime
  • Cut indexing and data-availability latency 5x, from 25 to 5 minutes, with concurrent, threaded execution in AWS Lambda, plus a config-driven framework to onboard new websites and PDFs
  • Prototyped an internal React demo UI with AI-assisted development to showcase the RAG content search service to product leaders, engineering teams and other lines of business
Show 3 moreShow less
  • Packaged the ML team's RAG code for an Amazon EKS pod, importing it without its full dependency set to shrink the image, then rolled the service out to production gradually at small traffic percentages
  • Tuned backend services for high-throughput read/write workloads with batching, parallel execution, query tuning and cache-friendly access patterns for low-latency reads
  • Gathered requirements from the product team and tested the RAG service against its evaluation documents, feeding needed changes, including prompts, back to the ML team
  • RAG
  • Generative AI
  • OpenSearch
  • Python
  • AWS
  • AWS Lambda
  • Concurrency
  • ETL

Apr – Oct 2024

Software Engineer

Infrastructure Migration: Terraform to CloudFormation

Moved the service from Terraform to the firm's internal environment-as-code framework, a wrapper over AWS CloudFormation templates

Zerodowntime during the migration
  • Set up service infra (Amazon EKS, Route 53, NLB/ALB, OpenSearch, Lambda) from the team's existing design; migrated it with the team lead from Terraform to an internal AWS CloudFormation framework, zero downtime
  • Deployed and operated Dockerized backend services on Kubernetes (Amazon EKS) with HPA autoscaling and resource quotas
  • Amazon EKS
  • AWS
  • Kubernetes
  • Amazon Route 53
  • Autoscaling
  • AWS CloudFormation
  • AWS Lambda
  • Docker

Jun 2023 – now

Software Engineer → SE II

Leadership, reliability & delivery

The work around the services: acting as team lead, running incident response, setting up release pipelines and chaos tests, and bringing new engineers up to speed.

Apr–Sep 2025acting team lead for 5 developers, a PM, an agile lead and QA
2junior engineers mentored and onboarded
  • Covered the team lead's duties (Apr–Sep 2025) for 5 developers, a PM, an agile lead and QA: led product discussions, refined stories, led SLA incident response and drove tech-debt work
  • Set up CI/CD pipelines in Spinnaker for every new service, with blue-green releases and automated functional tests across environments, enabling zero-downtime rollouts
  • Led resolution of 5–6 P3 and a couple of major P2 production incidents within SLA while covering the team lead, restoring key monitoring streams and metric pipelines with cross-functional teams
  • Ran Gremlin chaos tests that took down database and cluster nodes to validate failover of production services, and walked the chaos testing team through resiliency testing for our services
Show 4 moreShow less
  • Mentored a team of 8 first-year students from partner colleges in Figma prototyping through a 50-team JPMorgan Chase hackathon, guiding them to a final product pitch
  • Mentored and onboarded new joiners, including two junior engineers, through system walkthroughs, code reviews and learning plans, bringing one up to speed within ~2 months
  • Improved observability with Splunk, Datadog and Dynatrace for RCA tracking, with metrics broken down by region, namespace and availability zone
  • Tested the ML team's models and GenAI prompts for FAQ and Content Search on production, test and analytics data, alternating with leads; fed gaps back to the ML team to unblock API design
  • Mentoring
  • Observability
  • Automated Testing
  • Chaos Engineering
  • CI/CD
  • Datadog
  • Dynatrace
  • Figma

02 / Projects

Things I build after hours.

Daybook asks how many alerts fired, the answer is dictated, it reads the number back, and the summary shows resume-ready bullets

Side project · Oct 2026 · built for No Hands: The Wispr Challenge

Daybook

A voice-first work journal for software engineers. At the end of the day it talks with you for a few minutes, asks the follow-up a good manager would (“by how much?”), and files your answers as searchable memory: achievements with resume-ready bullets, the numbers behind them, and questions to close tomorrow.

I built it after sitting down to update my resume and finding that 21 of the 35 achievements in my notes had no measured result.

  • Gemini does the language work; plain Python decides everything else: what to ask next, when to wrap up, which numbers are allowed and how deltas are computed. It never invents a metric.
  • Each user’s memory is one SQLite file with full-text and vector search inside it, sealed with a recovery key only that user holds.
  • Serverless on AWS (Lambda, S3, CloudFront, EventBridge Scheduler) and scales to zero: about $0–0.10 a month in AWS cost for one person.
  • Speaks back with ElevenLabs, and exports the whole memory as Markdown written for any LLM, ready for a resume or an appraisal.
  • Python 3.12
  • FastAPI
  • React 19 PWA
  • TypeScript
  • Gemini
  • ElevenLabs
  • SQLite + sqlite-vec
  • AWS Lambda
  • S3
  • CloudFront
Try it liveWatch the 2-min demo
GitHubDev.toHashnode

The live demo opens a sandbox with two made-up weeks of reflections. No sign-in needed.

Hackathons · 2022 · Jan – Apr 2022

QuesGen

Web app that auto-generates self-assessment questions for students after an online class

Won Hackatra and Hack36, runner-up at Innerve 6.0 · ₹1,00,000 in prizes

  • Built QuesGen, a Flask and PyTorch app that auto-generates 4 question types (MCQ, paraphrase, boolean, FAQ) for student self-assessment; won 2 hackathons
  • Python
  • Flask
  • PyTorch
  • NLTK
GitHubDemoDevpost

More on GitHub

All repositories

Loading repositories from GitHub…

03 / Skills

Skills, counted by evidence.

The small number next to a skill is how many achievements in my vault use it. Faded ones I know but haven’t logged an achievement for yet.

AI / ML

  • Generative AI6
  • RAG6
  • Vector Search3
  • Context Engineering2
  • Embeddings2
  • LLM Evaluation2
  • AI Agents1
  • LLMs1
  • NLP1
  • Conversational AI
  • LLM-as-a-Judge

Languages

  • Python7
  • Java3
  • JavaScript1
  • C++
  • CSS
  • HTML
  • SQL

Frameworks

  • Spring Boot3
  • FastAPI2
  • Flask2
  • gRPC1
  • NLTK1
  • PyTorch1
  • React1

AWS

  • AWS6
  • Amazon EKS5
  • AWS Lambda3
  • Amazon Route 532
  • Amazon S31
  • AWS Step Functions1
  • Elastic Load Balancing1
  • Amazon ECS
  • Amazon VPC

DevOps & delivery

  • CI/CD2
  • Docker2
  • Kubernetes2
  • AWS CloudFormation1
  • Envoy1
  • Spinnaker1
  • Terraform1
  • Jenkins

Databases

  • OpenSearch6
  • Cassandra1
  • Redis1
  • FAISS

Data

  • ETL2
  • Web Scraping2
  • Kafkalearning
  • NumPy
  • Pandas

Practices

  • REST APIs5
  • AI-Assisted Development3
  • Mentoring3
  • Observability3
  • Automated Testing2
  • Concurrency2
  • Contract-First API Design2
  • Requirements Gathering2
  • Technical Leadership2
  • Autoscaling1
  • Caching1
  • Chaos Engineering1
  • Data Modeling1
  • Environment as Code1
  • Health Checks1
  • Infrastructure as Code1
  • PII Masking1
  • Query Optimization1
  • Secure API Design1
  • Distributed Systems

Tools

  • Datadog2
  • Dynatrace2
  • Splunk2
  • Claude1
  • Figma1
  • GitHub Copilot1
  • Gremlin1
  • Karate1
  • Protocol Buffers1

04 / Credentials

Certifications, awards & education.

Awards

Education

Bachelor of Engineering, Computer Science

Vellore Institute of Technology

Jun 2019 – Jul 2023 · CGPA 9.01/10

05 / Contact

Got a hard backend or AI problem? Let’s talk.

ayansadhukhan28@gmail.com