SMK
GitHubLinkedInCVContact
Contact

Hi, I'm

AI Engineer

I define what counts as working, then hold the model to it.

The setting changes from one project to the next; the way I take a problem apart and test it does not.

View projectsCV
years building AI systems
3+
Based in
Bengaluru, India

Featured Projects

Selected projects showing the problem, my role, system design, and outcome.

  1. Triage check-in kiosk for clinics · Research prototype2026

    CareFlow

    A patient with chest pain takes a ticket and sits down. Nobody knows what is wrong until their number is called.

    The problem
    Queues run in arrival order; danger signs do not. And a model left to sort the queue alone can miss the sign that matters most.
    My contribution
    So plain-code rules alert the nurse to danger signs as the patient checks in, and only then does an urgency model suggest a level for the nurse to confirm or override. Built the kiosk, phone and staff apps, the intake graph, the rules engine and the FHIR write-back.
    How it works
    • Kiosk-to-phone check-in in English or Arabic
    • Local Qwen model that turns typed answers into facts
    • Doctor briefing that shows where each fact came from
    Outcome
    Synthetic chest-pain, routine and sepsis journeys pass end to end, from kiosk check-in to a FHIR record.

  2. Regression testing for AI agents · Open source2026

    TraceBisect

    An agent that handled a case last week gets it wrong today. Both runs are logged, and the logs look almost the same.

    The problem
    The wrong answer is the symptom; the cause is one earlier step, like a tool called with new arguments. And nothing stops it coming back.
    My contribution
    So TraceBisect finds the first behavioural divergence between the two runs and exports a pytest test that re-runs the scenario and fails if tool arguments, branches, output or cost drift from the good run. Built the engine, the CLI and Studio, its FastAPI and Next.js dashboard.
    How it works
    • Trace format for LLM, tool, MCP and retrieval events
    • First-divergence detection with a type and severity
    • pytest export that re-runs the scenario in CI
    Outcome
    On the example agent, the exported test fails on the regressed run and passes on the good one.

  3. AI coding team workspace · Local developer preview2026

    AI Workforce

    Run three coding agents at once and you have three chat windows. Three briefs, three sets of changes, and no idea which ones passed.

    The problem
    The work is parallel; the accounting is not. Nothing tells you who owns which change, what is safe to merge, or whether anything was actually reviewed.
    My contribution
    So every agent works in its own copy of the repository, and nothing is marked verified until independent checks pass and a head review approves the exact candidate. I built the workspace, the orchestration engine, the MCP bridge and the review workflow.
    How it works
    • Task intake and department routing
    • Isolated candidate repositories, capacity-aware scheduling
    • Independent acceptance checks and review
    Outcome
    Demonstrated a tested, reviewed code candidate and parallel analysis across two projects.

  4. VS Code extension2026

    Grove — parallel AI development

    Two agents, one repository. They edit the same file, and you find out at merge time.

    The problem
    Isolation is the easy half — give each agent a worktree. Knowing they are about to collide, while there is still time to act, is the hard half.
    My contribution
    So overlap is detected while the work is still running and conflicts are predicted instead of discovered. Built the extension, the conflict checks and the guided merge.
    How it works
    • Teams of parallel Claude Code sessions
    • Isolated Git worktrees
    • Guided merge sequence with tests run between merges
    Outcome
    Open source on GitHub.

Experience

Professional Experience

  1. May 2023 — Present

    AI Engineer

    Odyssey Therapeia — Bengaluru, India

    Worked across retrieval, agent orchestration, conversational interfaces, and real-time systems.

    Selected contributions

    • Built a GPT-4o booking agent (Python, LangChain) to replace a fixed-flow booking chatbot. It picks from 14 tools per request, including RAG over Pinecone for doctor details.
    • Wrote the FastAPI REST API those tools call, with data on AWS: 5 Amazon DynamoDB tables and documents in Amazon S3.
    • Developed a Plivo voice agent for inbound and outbound calls (Deepgram, LLM, Sarvam) with barge-in, webhook signature checks, HMAC media tokens, Redis call state, Docker Compose and CI.
    • Designed a FastAPI and LangGraph hiring backend that parsed CVs with validation and retries, scored them against job descriptions and fed interview questions to an ElevenLabs voice agent.
    • Connected a device-side host to a remote specialist over two-way WebRTC audio and video, with a Python WebSocket signalling server, rooms, shareable meeting links and auto-reconnect.

  2. Feb 2023 — May 2023

    Software Developer Intern

    Odyssey Therapeia — Bengaluru, India

    Built neurotechnology tooling for research teams.

    Selected contributions

    • Built a React and Chart.js dashboard on the Neurosity Crown SDK to visualize live EEG data for neurological research.
    • Made live EEG data easier to inspect through interactive brain-activity visualizations.

Capabilities

Technical Skills

Tools and technologies used across my professional and project work.

  1. LLM & agent systems

    The reasoning layer

    • RAG
    • LangGraph
    • LangChain
    • MCP
    • Tool calling
    • Plivo

    Built retrieval and agent workflows with RAG, LangGraph and LangChain, and a phone voice agent on Plivo with Deepgram speech-to-text.

  2. Languages & backend

    The application layer

    • Python
    • TypeScript
    • SQL
    • FastAPI
    • WebSockets
    • REST APIs

    Developed Python services with FastAPI and WebSockets, and built a VS Code extension in TypeScript.

  3. Data & retrieval

    The context layer

    • DynamoDB
    • PostgreSQL
    • Redis
    • Pinecone
    • Embeddings
    • Vector search

    Modelled booking and record data across five DynamoDB tables with GSI-backed record queries, ran vector search on Pinecone, and used Redis for session state in Python services.

  4. Cloud & engineering workflow

    The tooling layer

    • Docker
    • Docker Compose
    • CI/CD
    • Git
    • AWS (DynamoDB, S3)
    • GCP

    Containerised and orchestrated multi-service stacks with Docker Compose, and built application-level integrations on AWS and Google Cloud.

About Me

Shebin Mohan K

Hi there! I'm Shebin, and I like making models useful. I build retrieval and agent systems, and the interesting problem is rarely the model — it's deciding what a correct answer looks like, then holding the system to it. I care about the parts that make that possible: grounding a workflow in real data, writing the evaluation before the feature, and leaving a clear path back to a person for the cases a model should not settle alone.

Currently, I'm an AI Engineer at Odyssey Therapeia in Bengaluru, where I build conversational and real-time systems for healthcare workflows.

  • Ground the system

    Connect each workflow to relevant data, clear sources, and the context needed for the task.

  • Evaluate the workflow

    Define success criteria and test the system against the problem it is meant to solve.

  • Design the fallback

    Keep a clear path for people when the model is uncertain or automation should stop.

years building AI systems
3+
co-authored publications
2

Education

Academic Background

HKBK Group of Institutions · Bengaluru, India

Bachelor of Engineering (BE), Computer Science

Aug 2019 — May 2023

Certifications

Licenses & Certifications

Google DeepMind In progress

Google DeepMind: AI Research Foundations

Research

Research & Publications

Co-authored research in AI-enabled cervical screening and image-processing surveillance.

International Journal of Gynecological Cancer, Feb 2026

VisiSpec — an AI-enabled vaginal speculum integrating coaxial illumination and intra-blade imaging for cervical screening

Shebin Mohan Kodumthara (co-author). International Journal of Gynecological Cancer 36(2):103307, Feb 2026 — conference abstract.

DOI: 10.1016/j.ijgc.2025.103307

International Journal of Novel Research and Development, June 2023

Automated Arm Robot for Intelligent Surveillance by Integrating Image Processing

Shebin Mohan K (co-author). International Journal of Novel Research and Development 8(6):b274–b279, June 2023 — final-year project.

Paper: IJNRD2306129

Contact

Let’s buildsomethinguseful.

I’m open to AI Engineer roles and collaborations, especially teams building LLM products on Python backends. If my experience fits your team or project, I’d be glad to talk.

shebinofficial14@gmail.com

Send a message