Software engineer building AI agents: production agents, agent memory and LLM evaluation.
M.S. Applied Machine Learning, University of Maryland (May 2027). Open to full-time roles from June 2027, San Francisco Bay Area or relocating.
3E · AI Engineer Intern, AI Platform · Jun to Aug 2026
Built the core of a multi-tenant agent-memory MCP service on AWS Lambda and DynamoDB (six operations, deny-by-default tenant access, versioned writes) and its deterministic spec-conformance harness. Found and fixed a warm-container credential leak; the fix landed on main. Benchmarked Hindsight, Cognee and Zep/Graphiti on LoCoMo and LongMemEval, and calibrated an LLM judge against blinded human graders: the rubric, not the judge model, was the accuracy lever.
Euler AI · Founding ML/Software Engineer · Mar to Jul 2025
Built a conversational shopping agent on FastAPI (intent routing, a reranker, PII guardrails) and a G-Eval pipeline that measured hallucination across the product.
EvoRank: LLM-Guided Evolution of Multi-Objective Learning-to-Rank Pipelines · code
First author, equal contribution · GenAIECommerce'26 workshop at RecSys 2026
An LLM evolves complete ranking pipelines, gated by a pre-spend headroom check, and its winners are judged on 59,902 held-out queries. All three runs beat an Optuna-tuned LambdaMART on relevance (best +0.0057 NDCG@10, p < 0.0001), and a one-shot late submission would have placed 20th of 340 on the original Kaggle test set.
Building Domain-Specific LLMs Faithful to the Islamic Worldview
Co-author · Muslims in ML workshop at NeurIPS 2023
Built the evaluation (BERTScore, embedding similarity) comparing prompting, RAG and GPT-3.5 fine-tuning, and presented the paper.
AI Resume Agent · live
Production AI agent on Gemini and FastAPI: 10 backend modules with graceful degradation, and five LLM-as-judge evaluators in Langfuse scoring every answer.
HVAC Copilot · YC × Cactus × Google DeepMind hackathon, 4-person team
Built the on-device agent runtime on Gemma 4 E4B: a six-tool function-calling surface with the on-device/online boundary enforced in the dispatcher, and a streaming parser for Gemma 4's tool-call formats. Profiled time to first token at 217 ms bare-model vs 3.9 to 4.5 s with ~935 tokens of system prompt and tool schemas, and traced it to CPU prefill.
FunctionGemma Router · 2nd place, Cactus × Google DeepMind hackathon · built all of it (three-person registration)
Hybrid on-device and cloud function calling (FunctionGemma-270M with Gemini 2.5 Flash Lite): 0.99 F1 at 548 ms average latency.
Top-K k-NN kernel for the Cerebras WSE-2 · CSL, take-home, code on request
Passes all 6 grader cases with 2.68× fewer cycles than the first correct version; a radix-sort variant was measured and rejected at +24%.
PathGuard · UMD × Ironsite hackathon, team lead
Architected the dual-pipeline system and built its on-device VLM narrator (Liquid LFM 2.5 via Cactus), whose scene-specific prompts drive the team's real-time hazard detection (Grounding DINO, SAM2, Depth Anything V2).
Open-source contributor to Mem0.