Skip to content
View vinsymoyya's full-sized avatar
🎯
Focusing
🎯
Focusing

Highlights

  • Pro

Block or report vinsymoyya

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
vinsymoyya/README.md

Build. Challenge. Verify. — AI engineering with inspectable evidence

Vinsy Moyya

AI/ML Engineer · LLM Evaluation & Terminal Benchmarking · Agentic AI · Scientific ML

I build AI workflows, investigate model failures, and contribute fixes to the tools behind them. My work connects hands-on engineering with reproducible evaluation: clear assumptions, useful edge cases, and evidence another person can inspect.

Portfolio · LinkedIn · Kaggle · ORCID

What I work on

  • LLM evaluation & benchmarking: model accuracy, reasoning, difficult cases, and structured feedback. My experience includes Terminal-Bench 2 and 3 tasks.
  • Agentic AI: agent orchestration, retrieval-augmented generation, and practical automation with LangChain, CrewAI, and n8n.
  • Open-source engineering: reproducible bug reports, scoped fixes, and regression checks across AI systems, SDKs, data tools, and cloud infrastructure.
  • Scientific ML: reconstruction, inverse problems, measurement uncertainty, and the limits of model generalization.

AI evaluation platforms

My work at Handshake AI focuses on AI evaluation and LLM benchmarking. I hold Platinum-tier recognition in the Handshake AI Fellowship.

I'm active across the following AI evaluation platforms:

Toptal Handshake AI Alignerr Crossing Hurdles
Deccan Experts AfterQuery Experts Turing Mercor

My evaluation focus is on making model behavior easier to assess: explicit criteria, reproducible tasks, reasoning checks, and actionable failure analysis.

Open-source contributions

I contribute across the following organizations and ecosystems. Each name links to my authored upstream pull requests.

Ecosystem Organizations
Cloud & developer tools AWS · Pydantic · Quix · Google · Tines
AI systems & infrastructure LlamaIndex / run-llama · ElevenLabs · Qdrant · SurrealDB · V7 Labs · BentoML · Baseten · ZenML · Doubleword
Scientific & quantum software Quantinuum
Evaluation & observability Evidently AI · Cleanlab · Braintrust

Selected engineering contributions

Area Example
Cloud tooling AWS Lambda Builders: validate uv exports while preserving lock options.
Agent workflows LlamaIndex / llama-agents: bind waiter timeouts to the registration that created them.
SDK behavior ElevenLabs Python: respect custom HTTP client timeouts.
Data handling Pydantic / aioaws: escape object keys in S3 multi-object deletion XML.
Evaluation Braintrust / AutoEvals: preserve explicit None overrides in partial scorers.
Go APIs Tines Go SDK: yield records before fetching the next page.

Explore all upstream pull requests

Research projects

When can an MRI reconstruction fit its measurements and still contain an unsupported structure?

An open benchmark of reconstruction errors and independent Fourier audits, with synthetic controls, open CNN weights, reproducible experiments, and an exploratory real-MRI extension. The work documents failure cases and the limits of audit sensitivity.

Real-acquisition report · Validation catalogue · Run the experiments

What do the measurements leave ambiguous, and which measurement should come next?

A toolkit for bounded linear inverse problems: conflicting admissible source fields, checked bounds on a requested quantity, and measurement selection. An independent checker uses exact rational arithmetic. The measured-acquisition study retains model-transfer failures alongside the certificates.

Methods · Acquisition study · Try the toolkit

How does reconstruction behave when acquisition conditions change?

A PyTorch toolkit for photoacoustic reconstruction, acquisition stress tests, and uncertainty experiments. Includes a synthetic GPU pilot, public checkpoints, reproducible evaluation, and measured-phantom acquisition diagnostics that expose geometry and calibration mismatches.

Acquisition failure study · Pilot report · Run the toolkit

These are exploratory research projects with unreviewed reports. Their results do not establish clinical validity or generalization across devices and acquisitions.

Engineering toolkit

Focus Tools
ML & evaluation Python · PyTorch · NumPy · pandas · scikit-learn
AI workflows LangChain · CrewAI · RAG · FastAPI · n8n
Software & infrastructure SQL · Docker · AWS · GCP · Java · Spring Boot · TypeScript

My background includes AI evaluation at Handshake AI, agentic AI training at Datavalley, ML research at Dyne, and Java full-stack development through TalentNet (Wipro).

Based in India · Open to AI/ML engineering, LLM evaluation, and research collaborations.

Let's connect on LinkedIn · Explore my portfolio

Pinned Loading

  1. k-sentinel k-sentinel Public

    Open MRI reconstruction audits with synthetic controls and a 100-participant real MRI pilot. Natural errors, calibrated audits, open weights and reproducible evidence.

    Python

  2. inverse-witness inverse-witness Public

    Checkable ambiguity witnesses and minimax next measurements for bounded inverse physics

    Python

  3. pa-recon-lab pa-recon-lab Public

    Reproducible photoacoustic reconstruction experiments with acquisition stress tests, uncertainty evaluation, and a Kaggle GPU workflow.

    Jupyter Notebook 1

  4. mindscribe mindscribe Public

    opensource extension

    JavaScript