Hi
I'm
I'm a Machine Learning Engineer working on AI security and LLM evaluation. At a health-tech startup I shipped production LLM and data systems: an agentic AI system that raised engagement about 25%, and streaming pipelines for 1M+ records/month. I publish security research (IEEE, Springer) and contribute fixes to the tools teams use to evaluate, red-team and serve models, at NVIDIA and EleutherAI.
01 / Selected Work
LLM evaluation and ML systems
Retrieves and reranks clinical notes, answers with citations, then checks every sentence with an NLI model and refuses when the evidence is not there. On 220 synthetic notes: 100% Recall@5 on 40 queries and 8 of 8 fabricated answers rejected.
clinical-notes-ragRuns both engines on the same model and hardware under five traffic patterns (uniform, bursty, ramp, cyclic, spike) and reports p50/p95/p99 latency, time to first token, throughput and cost per million tokens. A small-model router cut modeled cost 20.8%.
llm-serving-benchmarkScores a Kafka transaction stream with Spark Structured Streaming, watches features for drift with KS tests and retrains when drift turns critical. On 50,000 synthetic transactions: AUC 0.974, and retraining recovered AUC from 0.964 to 0.973 after an injected shift.
streaming-fraud-detectionA write-up of a silent bug: with shared-expert fusion on, a fused MoE LoRA Triton kernel reads the next adapter's weights instead of failing, so one adapter's requests run on another's weights. Includes the stride-arithmetic proof and a reproduction.
moe-lora-stride-aliasingBackend engineering
Go reverse proxy with per-client and per-route sliding-window limits enforced atomically in Redis Lua scripts. Exact under load: with a limit of 50 and 100 concurrent clients, 50 allowed and 50 rejected.
rate-limiter-gatewayDouble-entry ledger on FastAPI and Postgres with serializable transfers and idempotency keys. Zero balance drift after 1,000 concurrent transfers, zero duplicate charges under retries, zero overdrafts.
payments-ledger-apiPostgres-backed queue using FOR UPDATE SKIP LOCKED, with backoff retries, dead letters, heartbeats and crash recovery. No job lost or processed twice; 822 jobs/s with 16 workers.
distributed-job-queueIn-memory search engine in Go with an inverted index, BM25 ranking, and phrase and prefix queries, safe for parallel indexing and search. 2.8 ms median query on 10,000 documents.
fulltext-search-engineTools, products and teaching
A Python package on PyPI: one call on any dataframe returns the problem type, ranked models to try with a reason for each, data warnings and the charts worth looking at.
firstlookA cash register you talk to, for New York corner stores, in Spanish and English. The AI only turns speech into fields; plain code computes every dollar from an append-only ledger, with credit tracking, inventory and tax deadlines.
palapaA visual-first ML curriculum from zero to research, mapped as a 3D graph of what depends on what, and the open library of React Three Fiber and Manim visuals behind it, published on npm.
learn-mlHow to run an RTX 5060 Ti for ML inside a Proxmox virtual machine past NVIDIA's hypervisor check: what actually worked and what wasted time.
blackwell-proxmox-passthrough02 / Open Source
Benchmarks, fixes and scorers for the tools teams use to evaluate, red-team and serve models.
Contributed InfiniteBench, 11 long-context evaluation tasks (retrieval, code, math, novel QA and dialogue, in English and Chinese), to the open-source evaluation framework behind Hugging Face's Open LLM Leaderboard. 799 lines across 15 files, with scoring matched exactly to the official implementation.
Fixes to its red-teaming and evaluation packages: a judge failure on one item no longer throws away a whole red-team run, a target that matches an object, list or null no longer counts as an attack on an emptied input, and the trajectory evaluator reports on its 0 to 1 scale.
TensorRT-LLM: raw logprobs now come from the logits saved before biases and penalties rewrite them. TransformerEngine: MXFP8 and NVFP4 tensor ops, and misaligned grouped-GEMM operands. Megatron-Bridge: FP8 checkpoint import, RoPE on export, vocabulary padding. Dynamo: named chat templates.
03 / Experience
Seed-Stage Health-Tech Startup (NDA)
Seed-Stage Health-Tech Startup (NDA)
04 / Toolkit
05 / Research
Peer-reviewed work in AI security and trust: deepfake detection, ML for cybersecurity, and threats in AI-integrated cloud systems.
06 / Writing
Notes on building ML systems, evaluating frontier models, and lessons from production.
Why off-the-shelf benchmarks aren't enough, and how I built a custom harness with consistency scoring, calibration, and chain-of-thought quality analysis.
Read rag & agentsArchitecture decisions, FAISS optimization, and why agent evaluation is harder than agent building. From processing 2.3M figures, 890K tables, and 410K equations.
Read research toolingExtending TRIBE v2 with streaming inference, modality attribution, brain-alignment benchmarking, and cognitive-load scoring, built into a released toolkit.
Read07 / Education
Stevens Institute of Technology
Hoboken, NJSwami Rama Himalayan University
Dehradun, India08 / Contact
Graduating December 2026 and starting January 2027, with three years of US work authorization (STEM OPT). Open to ML engineering, applied AI and research-engineering roles, especially in AI security and LLM evaluation. Based in Hoboken, NJ (New York area); open to relocating anywhere in the US.