lokesh sandhu · vancouver, bc · --:-- PT
// engineering with intent



Computer science at UBC, specializing in AI. Currently shipping AI-powered clinical documentation at ScribeFlow and doing research at UBC Computer Science.

Routing under constraints, systems under load, pipelines that stay traceable. Graduating December 2027, looking for internships.

/log
// experience: select an entry for the full notes
  • Batched 25 AI question embeddings per request, eliminating 96% of embedding-adapter calls across 1,000 synthetic questions; cosine scores matched the reference within 1e-12 on every comparison.
  • Compressed revision storage 57.2% versus full raw copies across 100 synthetic revisions with content-addressed deduplication, manifest overhead included; restored all 400 stored files with zero byte mismatches.
  • Validated canonical SHA-256 digests against 3,000 synthetic content, dependency, and version changes with zero missed changes, and zero false positives across 1,000 equivalent inputs with reordered JSON keys.
  • Built the transcript-to-note backend with queued AI generation jobs, schema validation, and clinician review: 99% processing completion across 1,000 synthetic patient encounters with injected job failures.
  • Cut p95 note-generation latency 40% (20s → 12s) and cost per completed note 25% by profiling workloads, shortening prompts, and bounding concurrent API requests, without losing held-out evaluation quality.
  • Developed a 200-case clinical-note evaluation suite detecting omitted facts, unsupported claims, and malformed structured outputs, for repeatable comparisons across model and prompt changes.
  • Gated CI on those 3 regression classes, with job-level failure tracing and rollback checks to diagnose and recover failed releases.
  • Shortened median source-verification time 30% (10 → 7 min) by linking note sections to transcript excerpts, so reviewers can trace medical details before clinical approval.
  • Evaluated trading signals against historical market data with Python backtests, exploratory analysis, and statistical tests comparing strategy behaviour across variants.
  • Compared portfolio risk and strategy performance across research iterations, documenting assumptions and tradeoffs so reviewers could reproduce the reasoning behind each result.
  • Refined research hypotheses through recurring reviews of backtest outputs, improving consistency across strategy comparisons.
  • Supported 35+ residents by triaging concerns, mediating conflicts, and connecting students with campus resources or staff when they needed more support.
  • Maintained continuity on sensitive cases by recording incidents, follow-ups, and recurring concerns while following confidentiality and residence policies.
  • Helped staff intervene earlier by sharing recurring community concerns and case updates, and planned community programs and orientation activities with Residence Life.
  • Built Stripe payment verification and reconciliation with authenticated webhooks and idempotent order state transitions: consistent payment records with no duplicate fulfillment across 30 simulated payment-failure scenarios.
  • Developed and ran the storefront selling dorm starter kits to incoming UBC students (listings, pricing, checkout), iterating on features and UX directly with the founder.
  • Senior Finance Coordinator, Punjab Students Association: managed a $10K+ programming budget, tracking expenses, reimbursements, and documentation across student events.
  • Broadcast Executive, UBC Esports: produced YouTube videos and reels, configured OBS scenes, overlays, and capture hardware, and troubleshot live productions in person and remotely.
  • VP Finance, Multi Faith Club: budget, reimbursements, and spending-limit communication for recurring programs since 2024.
/work
// select an entry to expand: six files, no filler

A router coordinating warehouse robots: cached shortest distances, time-space traffic reservations, capacity-aware delivery batching. Delivered 24/24 inventory pods where the baseline managed 16/24; mean simulator score 58.66 → 83.99 with bounded-horizon routing and memoized task assignment.

1st place, Amazon Robotics Hackathon 2026: 4,766.62 points.

python · optimization · simulation · source ↗

Rebuilds order books for 8,892 symbols from a real 268.7M-message trading day at 9.94M msgs/s, zero mismatches against a reference book across 1,340 checkpoints. Throughput 1.96×, p99 latency 865 → 417 ns by replacing std maps with a flat open-addressing table and sorted-vector price levels (bootstrap confidence intervals, 30 runs per design).

93 unit tests, an independent BBO oracle, fuzzing and sanitizers in CI. libFuzzer caught a sentinel-key collision, ASan a use-after-erase; both are regression tests now.

c++20 · google benchmark · libfuzzer · asan/ubsan · source ↗

An AI itinerary planner. Cut planning API usage 60% across 100 replay sessions with request deduplication and a 60-second, 20-entry cache, validated by five test controls covering expiry, invalidation, retries, isolation, and eviction.

Cut total route distance 15.9% across 1,000 synthetic six-stop itineraries by comparing 12 candidate stop sequences through bounded relocation search, instead of brute-forcing all 720 permutations.

typescript · next.js · react · gemini · postgresql · source ↗

Led development of client registration and appointment scheduling around a real food bank's workflows. Normalized schema, JWT role-based REST APIs, AES-256-GCM field encryption bound to table-and-column metadata. On synthetic stress tests: zero overbooked slots across 1,000 concurrent booking attempts, zero partial reservations after 100 injected database failures.

node.js · express · postgresql · jwt · aes-256-gcm

Turns a semester of scattered Canvas files into structured, reviewable study material: ingestion into JSON keyed by course → lecture → topic → source, three unified input paths (slides, voice, manual), every intermediate representation validated so AI generation stays traceable instead of becoming a black box.

python · hugging face · canvas api · json schema

Tactical Route Planner: drone-imagery threats become GPS constraints, routes recompute live over real road networks; React + FastAPI + NetworkX + Gemini Vision, built end-to-end in a 4-hour hackathon.

Canvas CLI: browser-free Canvas from the terminal, incremental sync, SQLite-backed.

Apple Reminders MCP: lets AI assistants manage local Apple Reminders through macOS EventKit.

UBC Room Directory: full-stack classroom comparison with Google Maps distances (CPSC 310).

/recognition
2026 · 1ST PLACE · AMAZON ROBOTICS HACKATHON · 4,766.62 PTS, 24/24 PODS DELIVERED
2026 · SHIPPED IN 4 HOURS · TACTICAL ROUTE PLANNER, HACKATHON BUILD
2026 · 9.94M MSGS/S · NASDAQ ITCH RECONSTRUCTOR, 268.7M-MESSAGE TRADING DAY
/toolkit
// what the files above are written in
languagesC++, Python, Java, Go, Rust, TypeScript, JavaScript, SQL, R, Bash frameworksReact, Node.js, FastAPI, Django, Flask, PyTorch, TensorFlow, pandas testingGoogleTest, Google Benchmark, libFuzzer, ASan/UBSan, pytest, Jest infraAWS (S3, SQS, IAM), Kubernetes, Docker, GitHub Actions, GitLab CI, Linux
/contact
Come say hi.

I'm looking for internships where the problems are messy and the systems have to keep working anyway. Drop me a line.

lokeshsandhu4605@gmail.com
NORMAL ~/lokesh/portfolio.md : for commands · try :theme, :work, :help --:-- 0%