Accelerated CS @ Boston University, graduating a year early. I build production systems as the sole engineer — a degree planner used by 1,000+ students, and a nonprofit's entire ~300-endpoint platform. Alongside that, I contribute to the LLM inference stack (vLLM, llm-d, KServe, InstructLab): my niche there is finding the paths a new feature quietly fails on.
Previously — AI & Software Engineer Intern @ HireBuddy · Lead Software Engineer @ Hack4Impact BU
I contribute to vLLM, llm-d (a CNCF sandbox distributed LLM inference stack founded by Red Hat, Google Cloud, IBM Research, CoreWeave and NVIDIA), KServe, Grafana, Redis, Keras, and Red Hat's InstructLab. Also Prisma, MLflow, Datadog, SGLang and the Model Context Protocol registry. Same method everywhere: reproduce the failure with executed evidence before proposing the fix.
67 pull requests across 24 repositories in 17 organizations |
26 merged — including a fix in Grafana core shipping in 13.3.0, plus vLLM, node-redis, Keras and KServe |
9 defects found and filed against a newly released no-Kubernetes deployment path, 6 reproduced live with tracebacks |
19 reviews on other contributors' pull requests across 10 repositories |
What I am actually useful for: running a brand-new feature on a machine its maintainers do not have, a clean macOS laptop with no cluster and no GPU, and finding the paths where it reports success while doing nothing. That produced nine issues against one feature, including a container name collision that let one deployment's teardown destroy a sibling deployment's running containers.
Reviewing counts too. On someone else's pull request I found that a proposed dependency pin would fail the project's own version comparison, so every install would re-run the installer and silently overwrite the user's kubectl binary. Proven by running the project's own function against the proposed value.
|
Autonomous banks, central banks & payment rails (RTGS/ACH/SWIFT/CLS/DTCC), systemic-risk analysis, crisis replay & a hand-written regulatory DSL.
|
Degree planning for 1,000+ BU students across nearly 10,000 courses. Requirement-validation engine, vector-embedding semantic recommender replacing a quota-limited LLM at zero marginal cost, Redis layer at ~81% hit rate over 660k+ lookups.
|
|
The production system of record behind my CTO role — member lifecycle, payment-gated resources, evaluation forms & a film festival, with per-chapter tenant isolation.
|
Built the AI candidate-sourcing engine (Elasticsearch-DSL over CoreSignal) — tripled candidates surfaced per search (100 → 300) — and co-built a GPT-4o-mini pre-screening engine scoring across 10 evidence-backed dimensions.
|
OCR confidence-preservation layer for a specimen-digitization ML platform — a versioned per-field certainty envelope surfacing Azure vision-OCR to reviewers, plus SSRF + authorization hardening. Client-owned (BU Spark!) — no public repo |
Owned the SE-Chem badge-authoring wizard end-to-end — decomposed an 843-line page into components, killed a stale-closure bug, and batched Client-owned (BU Spark!) — no public repo |
%%{init: {'theme':'base','themeVariables':{'primaryColor':'#171F33','primaryTextColor':'#E6EDF3','primaryBorderColor':'#14B8A6','lineColor':'#10B981','fontSize':'14px'}}}%%
flowchart LR
UI["Frontend\nReact · Next.js · Three.js"] --> API["APIs\nFastAPI · Flask · Express · DRF"]
API --> AI["AI Layer\nLangGraph · MCP · embeddings · TF-IDF retrieval"]
API --> DATA[("Data\nPostgres · Redis · Elasticsearch")]
AI --> DATA
API --> INFRA["Infra\nDocker · AWS · Vercel · Railway"]
DATA --> INFRA
