Soham BapatSenior MLOps & AI Engineer
I build AI systems that run unattended — monitoring, auditing, and healing themselves.
I keep production machine learning and LLM systems dependable at a regulated bank — model scoring, retrieval-augmented assistants, pipelines, monitoring, guardrails.
ReadWhat broke in my autonomous trading agents, and what I changed- Current
- MLOps Automation Senior Engineer, Huntington National Bank
- Experience
- 4+ years
- Location
- Columbus, Ohio — remote preferred, relocation open
- Authorization
- U.S. Citizen
- Seeking
- MLOps · LLMOps · AI Engineering · ML Platform
Huntington National Bank
MLOps Automation Senior Engineer
Feb 2025 — Present
Full-time · Contract-to-hire · Columbus, OH
I own scoring, monitoring, and auditing for the bank's marketing models, which generate an estimated $100M in annual revenue, and I plan the project's development roadmap. Alongside it I build the bank's internal generative-AI systems, and run pipeline and orchestration work across the same AWS stack.
Scoring, monitoring, and auditing
I own scoring for 57 models in production
Scored monthly and published to the enterprise data lake, where they drive the bank's marketing and campaign decisioning.
Scores and audit reports land on schedule
I am responsible for delivering both on time. In a regulated environment the audit trail travels with the score, so the reports are part of the deliverable rather than a follow-up.
Monitoring and auditing the models themselves
Drift monitoring, telemetry, and automated report delivery on Terraform-provisioned AWS infrastructure, released through Azure DevOps CI/CD.
I plan the development roadmap
Which models onboard next, what the release cadence can absorb, which manual steps to automate, and what to rebuild before it turns into a bottleneck.
Downstream consumers are part of the job
Other projects read these scores, so I coordinate schema and schedule changes with those teams ahead of time and give them a window to adapt.
Internal AI assistants, and agent tooling
Retrieval-augmented assistants, one serving hundreds of colleagues
I build the RAG pipelines behind the bank's internal assistants. On the largest, staff ask questions in plain language against internal policy documentation and get a synthesized answer with clickable citations back to the passage it came from. The citations let a reader verify the answer against the source, which is what makes it usable for policy questions.
Ingestion: scheduled chunking and embedding into a vector store
Step Functions and EventBridge drive the pipeline on a schedule. Documents are pulled from the internal source, chunked, embedded with Cohere Embed v4, and indexed into OpenSearch Serverless.
Query path: agent in Lambda, Claude Haiku in Bedrock, guardrails in front
A Strands agent running in Lambda retrieves the relevant chunks and passes them to Claude Haiku on Bedrock through a content guardrail before the answer is returned. Haiku keeps per-query cost and latency low enough for interactive use.
Five Lambdas, each with a single job
The service is split into five Lambdas — agent, document pull, indexer, feedback capture, and query logging — so they scale, fail, and get debugged independently. React 19 and Vite on the front, FastAPI behind it, SSO through Keycloak, deployed on Shakudo.
Agentic ML-ops tooling on Bedrock and MCP servers
Internal automation that runs ML-ops workflows end to end, built with n8n and Claude Code.
Pipelines, orchestration, and ETL
A 7.5M-record customer pipeline
EventBridge, Step Functions, Glue, and SageMaker Pipelines, streamed to Kafka with Splunk log monitoring.
The same problems across related projects
Scheduling and retries, catching runs that finish without producing output, and Glue ETL with PySpark and Boto3 over Athena and data lake tables.
Presenting and handing off
Solo speaker at the company Data Expo
Presented the marketing models work, after a poster at the same event the year before. Also built internal AI tooling at the company hackathon.
Rebuilt handoffs with data science, support, and offshore
Made ownership explicit at each team boundary instead of assumed, and run knowledge-transfer sessions that keep delivery continuous across time zones.
- AWS
- Bedrock
- OpenSearch Serverless
- Cohere Embed v4
- Strands agents
- Lambda
- Step Functions
- EventBridge
- FastAPI
- React 19
- Keycloak
- Shakudo
- Terraform
- SageMaker
- Glue
- Kafka
- Azure DevOps
- Splunk
- Python
- PySpark
Autonomous LLM trading agents
Personal project — designed, built, and operated solo
May 2025 — Present
Self-hosted on Raspberry Pi hardware over a private Tailscale network
Read the full write-up, including what went wrong →Two LLM agents that research, decide, and trade with no human in the loop, on hardware in my apartment rather than in a cloud account. Most of the work went into the operational side: getting models, hardware, and a home network to behave predictably together.
On my own hardware, on my own network
Two Raspberry Pis running in my apartment, reachable only over a private Tailscale mesh — no public ingress, no cloud account, no managed control plane. I own the boards, the OS, the network, the deploy path, and the recovery story.
Tiered Claude routing by task complexity
Haiku for triage, Sonnet for analysis, Opus for decisions that warrant it. Routing by complexity keeps cost and latency predictable compared with sending everything to one model.
Auto-reverting experiment loop over an 8-year backtest
Agents evaluate their own performance against the backtest and automatically revert changes that measure worse than what they replaced.
Multi-level circuit breaker
Three levels with distinct triggers, each halting a different amount: one position, the session, or all trading.
Nothing about running it requires me at a keyboard
systemd units restart on failure, the two Pis health-check each other, and backup and recovery restore state without intervention. It has come back on its own through power blips and network drops.
- Python
- Anthropic Claude API
- Raspberry Pi
- Tailscale
- Linux
- systemd
- SQLite
- Backtesting
Vosyn.ai
Cloud / AI Platform Engineer
Jun 2024 — Feb 2025
Contract · Deputy team lead
Built the speech-translation pipeline behind the product, and led the platform migration from AWS to GCP underneath it.
An audio translation pipeline, stage by stage
Incoming video is demuxed into separate audio and video tracks. The audio is transcribed with Whisper, the transcript is machine-translated into the target language, and the result is re-voiced with text-to-speech before being remuxed against the original video.
Each stage its own service
Transcription, translation, and synthesis run as separate serverless services rather than one long job, so each scales on its own and a slow stage does not hold the others open.
Lead engineer on the AWS to GCP migration
Terraform and GKE, re-architecting task-based Docker execution into a service-based ECS architecture.
CI/CD for serverless microservices across both clouds
GitHub Actions automating workflows on Lambda and Cloud Run.
SageMaker pipeline migrated to Vertex AI
Part of the platform transition. Led a team of five deploying Dockerized Flask endpoints across ECS, Lambda, EC2, and ECR.
- Whisper
- Speech-to-text
- Machine translation
- Text-to-speech
- GCP
- GKE
- Vertex AI
- Cloud Run
- Terraform
- GitHub Actions
- ECS
- Docker
TELUS International
Data Analyst / GenAI Evaluation
Dec 2022 — May 2023
Contract · Remote
Evaluation data for generative AI models — the labeling side of model quality, before I was building the models.
Human-aligned datasets for generative AI
Evaluated and rated query results from major map services, producing labeled data used to fine-tune and evaluate generative AI models.
Consistency is what makes a dataset trainable
Ratings have to agree with each other before a model can learn anything from them, and I held roughly 100% accuracy on collection and entry across the engagement. It is the same concern that later made me build a backtest harness before trusting an agent's own scoring.
- GenAI evaluation
- Human-aligned data
- Labeling quality
- Qualitative research
How I work
The part that is not code
I have worked both ways: on a team with a lot of handoffs, and alone on systems nobody else touches.
Contributing and leading
- At Huntington the models come from data science and the delivery is mine — scoring, monitoring, and auditing, plus the release schedule both sides plan around.
- On the RAG work I implement to a solution architect's design, and I rewrote the handoff process with data science and support so each boundary names who owns what.
- Run knowledge-transfer sessions with an offshore team across a large time difference, so work continues when we are not both online.
- Deputy team lead at Vosyn: delegated a five-engineer workstream across ECS, Lambda, EC2, and ECR, and was lead engineer on the AWS to GCP migration across cloud, ML, frontend, and QA.
Unsupervised, for months at a time
- Two agents running unattended since May 2025 on my own hardware — no team, no on-call rotation, nobody to escalate to.
- Diagnosed five separate faults in one postmortem, then built, tested, deployed, and checksum-verified every one of them across both machines the same day.
- Automate my own toil first: systemd restarts, peer health checks between the two machines, and backup and recovery, so keeping it running costs me no time.
- Model routing by task complexity keeps cost predictable, and every change is gated behind an eight-year backtest that reverts its own regressions.
Earlier roles
- Technimark LLCDatabase Management InternPower BI · 100+ DAX calculations2023
- Manhattan AssociatesSoftware AnalystSupply chain and POS platforms · Jenkins CI/CD2021 — 2022
Education and certification
- M.S. Computer ScienceNorth Carolina State University2024
- B.S. Computer ScienceKennesaw State University · Honors2021
- Google Cloud Fundamentals for AWS ProfessionalsGoogle CloudVerify badge ↗2024
- AWS Certified Cloud PractitionerAmazon Web Services · Valid through Jan 2029Certificate ↗2023
Tools and platforms
AI and LLM systems
- RAG pipelines
- Amazon Bedrock
- OpenSearch Serverless
- Cohere Embed v4
- Vector search
- Strands agents
- Anthropic Claude API
- Whisper
- Speech-to-text and TTS
- MCP servers
- Agentic workflows
- Model routing
- Prompt engineering
- Bedrock Guardrails
- Evaluation
- n8n
ML platform and MLOps
- SageMaker AI
- SageMaker Pipelines
- Vertex AI
- Model scoring
- Drift monitoring
- CI/CD for models
Cloud and infrastructure
- AWS
- GCP
- Terraform
- Docker
- Kubernetes
- Lambda
- Glue
- Step Functions
- EventBridge
- ECS
- ECR
- Athena
- IAM
Data and languages
- Python
- SQL
- Java
- Apache Spark
- Kafka
- Pandas
- PySpark
- BigQuery
- Power BI
- Splunk
Application layer
- FastAPI
- React 19
- Vite
- TypeScript
- Keycloak SSO
- Shakudo
- REST APIs
Hiring for production AI? Let's talk.
Open to MLOps, LLMOps, AI engineering, and ML platform roles. Remote preferred, relocation open. U.S. Citizen, based in Columbus, Ohio.