Life Changelog.

887% revenue acceleration through repeatable forward-deployed engagement model.

• → FDE feedback loop ->

Joined Benmore as Employee #2 — Forward Deployed Strategist & Engineer. Embedded into client engineering teams, scoped systems end-to-end, and shipped production code from day one.

• Onboarded the first ten clients across SaaS, healthcare, and consumer verticals.
• Authored the Benmore Foundry CLI — internal orchestration layer for SMB AI consulting engagements.
• Cross-stack: Stripe, Django, Next.js, FastAPI, React Native, plus Claude Code skill systems.

PythonTyperDjangoFastAPINext.jsReact NativeStripeClaude Code

• → Experience & work ->
• → Why FDE ->

Designed and trained the first 3D self-supervised vision transformer for subcellular structure prediction from cryo-electron tomograms — a result that beat a prior approach led by a Nobel laureate.

• Adapted DINO-style self-supervised contrastive learning into 3D — student/teacher ViTs over volumetric tomograms.
• Pretrained on unannotated tomograms; fine-tuned on a tiny labeled set for vesicle / organelle classification.
• Beat the prior SOTA, including the Nobel-laureate-led approach, on downstream evaluation.
• Released as a BioRxiv preprint, first-author.

PyTorchDINO3D ViTSelf-Supervised Learningcryo-ETFSDP

• → BioArxiv preprint
• → SpatialDINO lessons ->
• → View papers ->

ML researcher at Harvard Medical School applying 3D vision transformers to cryo-electron tomography. The lab images subcellular structures at near-atomic resolution; my job was to make the resulting volumes interpretable at scale.

• Trained on multi-node DGX clusters: A100 / H100, NVLink intra-node, Infiniband inter-node, RAID + NVMe storage tier.
• PyTorch FSDP with bf16 mixed precision and activation checkpointing to fit large 3D ViTs.
• Diagnosed a Rendezvous (RDZV) backend issue affecting Infiniband multi-node training — filed PyTorch issue #144779.

PyTorchFSDPCUDAInfinibandNCCLDGXcryo-ET

• → View research ->
• → PyTorch #144779
• → Kirchhausen Lab profile
• → Harvard ID (PDF)

BA in Math & CS, MS in CS, completed in four years.

• → BA degree (PDF)
• → MS degree (PDF)

BU Master's thesis: dynamic checkpointing for Apache Flink. Static checkpoint intervals are a tax in idle periods and a stall during bursts; I built a controller that adapted cadence from live backpressure signals.

• Instrumented Flink JobManager to surface per-operator backpressure ratios as a control signal.
• Adaptive checkpoint cadence: shorten when load is low, lengthen under sustained backpressure to avoid amplifying stalls.
• RocksDB state backend benchmarked against in-memory; quantified write-amplification tradeoffs.
• Validated on the NEXMARK streaming benchmark — measured tail-latency wins on bursty workloads.

JavaApache FlinkRocksDBNEXMARKJVM

Six classes spanning programming-language theory (CS 320 — OCaml interpreter from scratch), optimization (DS 522 — SGD variants & implicit SGD), cloud (CS 561 — PageRank on a synthetic mini-internet), graduate algorithms (CS 630 — Gale-Shapley average-case analysis over 331,776 instances, reservoir sampling), mathematical statistics (MA 582), and a liberal-arts elective (HI 151).

• CS 320: hand-written lexer, parser, and stack-machine evaluator for a BNF language in OCaml.
• DS 522: Adam / AMSGrad / RMSProp comparison, plus article evaluations of Reddi 2018 and Toulis 2016.
• CS 561: built a generated-HTML mini-internet and ran PageRank locally vs. on GCP.
• CS 630: 331,776-instance enumeration of Gale-Shapley to study average-case behavior.
• MA 582: rigorous graduate inference — MLE, sufficient statistics, MGF, asymptotics.

OCamlPythonRGCPOptimizationStatistics

• → CS 320 OCaml interpreter
• → DS 522 SGD variants
• → CS 561 PageRank
• → CS 630 Gale-Shapley
• → MA 582 Mathematical Statistics
• → All coursework

My first production-grade Rust. Built an Excel-side marketplace + a server-side ingestion pipeline that converted unstructured data (CSVs, JSON dumps, free-form Excel) into structured records flowing through NATS JetStreams. Spent the first two weeks deep in The Rust Book — the borrow checker forces you to internalize ownership, lifetimes, and Send/Sync before you can ship anything async.

• tokio + async-trait for concurrent I/O across hundreds of NATS subjects.
• NATS JetStream for at-least-once delivery, persistent streams, and replay.
• Merkle-tree POC (SHA-256 + canonical JSON hashing + sorted pairwise concat) for tamper-evident sync of records across the pipeline.
• Excel side: JavaScript Office Add-in scaffolded with Yeoman (`yo generator-office`).
• Generated TLS dev certificates with `office-addin-dev-certs`, trusted them in macOS Keychain, wired their paths into `nats.conf` so the add-in published over TLS.
• Borrow checker lessons: `&str` vs `String`, ownership over defensive `.clone()`, Send + 'static for futures.

Rusttokioasync-traitserdesha2NATS JetStreamYeomanJavaScriptOffice Add-ins

• → merkle_tree (Rust POC)
• → Merkle tree write-up
• → excel_connector (Yeoman + NATS)
• → The Rust Book

Distributed Systems (CS 350) — Raft consensus from scratch in Go, plus a MapReduce coordinator/worker. OOP & Design Patterns (CS 611) — Monsters & Heroes turn-based RPG plus a final Java Swing trading platform.

• CS 350: full Raft state machine with leader election, log replication, and snapshot install RPC.
• CS 350: MapReduce coordinator + worker with plugin-loaded map/reduce functions.
• CS 611: Monsters & Heroes — battle, market, inventory, 2D grid world.
• CS 611: group capstone trading platform with Java Swing GUI and singleton persistence.

GoRaftJavaOOPDesign Patterns

• → CS 350 deep dive
• → CS 611 Monsters & Heroes
• → CS 611 Trading System
• → All coursework

Battery taught me VC end-to-end. Worked under Dallin Bills sourcing early-stage B2B software - learned the difference between vertical and horizontal SaaS, fintech investment theses (payments, infra, embedded, neobanks), and how Battery sized markets across verticals. Got fluent in the financial bar: Rule of 40, ARR growth vs. burn multiples, gross retention vs. logo churn vs. net dollar retention, magic number, and CAC payback. Sourced 3 deals to partner-meeting stage, including CarNow.

• → Battery Ventures
• → CarNow (Battery portfolio)
• → Michael Brown - GP
• → Dallin Bills (LinkedIn)

"Ultrafast Two Dimensional Infrared Spectroscopy of Supercritical Fluids: Energy Relaxation and Local Critical Slowing Effects." NSF-funded UROP under Lawrence Ziegler (CAS Chemistry).

• → BU CS profile
• → UROP Symposium 2021 brochure (PDF)
• → UROP 2020-2021 awarded students

Left Chandigarh for Boston University as an undergraduate.

Two years of E&M, particle physics, organic and physical chemistry, and optics - the standard JEE Advanced gauntlet for kids who wanted to do physics in India. Also sat AP Calculus, AP Physics C: Mechanics, and AP Physics C: E&M.