Applied AI / ML Systems Engineer working across LLM evaluation, model infrastructure, inference, retrieval, and reliability.

Currently a Founding Engineer at Uniiq.ai, where I work across AI workflows, backend systems, reliability, security, testing, infrastructure, and product performance.

My public engineering work is split between an AI Systems Lab — where I implement the systems underneath modern AI from first principles — and upstream open-source work in established AI/ML codebases.

I contribute reliability and correctness fixes to open-source AI/ML infrastructure.

Merged/upstream work across areas including:

• cache and filesystem reliability
• CLI parsing semantics
• evaluation and benchmark correctness
• few-shot/evaluation behavior
• distributed result handling

A cache-path fix I contributed in PR #4047 was merged upstream and shipped in v0.4.13.

Bug investigation across agent/workflow execution, Editor state, evaluation/scorer caching, processor graphs, provider configuration, and CLI behavior.

Several reported issues progressed into upstream resolution / pending-release states.

Active upstream work across AI gateway and provider reliability:

• URL/path and browser-origin boundaries
• shell-safe configuration output
• stable API error contracts
• provider pagination
• package typing
• credential and recovery behavior

Active bug investigation and fixes across:

• terminal/WebSocket behavior
• Unicode and encoding boundaries
• filesystem portability
• Git/repository semantics
• pagination and cache identity
• authentication and URL handling

Earlier engineering contributions include work in the p5.js Web Editor, VulnerableCode, and PyNN, alongside scientific-ML work in the ML4SCI ecosystem.

See OPEN_SOURCE.md for the contribution ledger and current upstream status of individual artifacts.

OpenAI-compatible multi-provider LLM gateway with routing, failover, distributed circuit breaking, rate limiting, semantic caching, cost tracking, and observability.

LLaMA-style distributed training from PyTorch primitives using FSDP/ZeRO-3 and tensor parallelism, with sharded checkpoints, deterministic resume, correctness tests, and profiling.

LLM evaluation harness covering dataset validation, provider abstraction, deterministic and model-based scorers, concurrent execution, crash-safe artifacts, reporting, and regression workflows.

PyTorch quantization library implementing INT8, FP4, and NF4 from first principles with calibration, serialization, numerical validation, and benchmarks.

Sandboxed execution service for untrusted code using isolated containers, namespaces, seccomp, cgroups, filesystem/network restrictions, and resource controls.

Behavioral debugging for neural networks across checkpoints using representation analysis, interventions, training-data attribution, and counterfactual evidence.

I maintain a broader set of from-scratch/reference implementations covering:

Model internals — Transformers, SSM/Mamba, diffusion
Training & alignment — distributed training, PEFT, DPO
Evaluation & debugging — EvalForge, Neural Bisect
Inference & efficiency — quantization, ONNX serving, decoding
Retrieval & context — vector search, Graph RAG, prompt caching
Infrastructure & reliability — gateways, guardrails, secure execution

The goal is to understand and validate the abstractions underneath modern AI systems rather than only compose high-level frameworks.

Worked on privacy-preserving deep learning through an AI Pioneers open-source research collaboration mentored by Yale researchers Haris Smajlović and Claus Horn.

The work focused on adapting CNN components for Secure Multi-Party Computation in the Sequre ecosystem and building an end-to-end ChestMNIST training workflow.

LLM evaluation · AI infrastructure · model inference · ML systems · retrieval · reliability · privacy-preserving ML

Portfolio: https://div90.vercel.app/