Self-hosted interview prep platform targeting quantitative developer roles. The execution backend spawns compiler subprocesses (g++ with selectable C++ standard up to C++20, python3), capturing stdout against expected output or running assertion-based test drivers with stderr surfaced as structured diagnostics.

Scheduling engine implements FSRS-4.5 from specification — tracking per-item stability, difficulty, and discrete queue state (new → learning → review → relearning → leech) with the full forgetting-curve model driving next-review intervals. A dependency graph maps C++ topics to learncpp.com chapters with quant-relevance weights; the daily scheduler traverses the graph interleaving cards across topic clusters with per-topic caps and deduplication. No external tooling; fully self-contained and offline.

• Bids and asks stored in sorted std::maps (descending/ascending), each price level holding a std::list<Order> to preserve FIFO queue semantics
• Cancel and modify are O(1) via an unordered_map of iterators into per-level lists — no full scan. Modify re-queues only on volume increase, preserving priority for partial fills

From-scratch reimplementations of core STL containers against the standard's complexity and iterator-invalidation guarantees: std::vector (dynamic array, amortised O(1) push_back, documented reallocation contract), std::list (doubly-linked, O(1) splice, cache-locality cost versus contiguous alternatives), std::priority_queue (binary heap over vector backing, sift-up/sift-down with index arithmetic), std::unordered_map (separate chaining, configurable load factor with full bucket redistribution on rehash).

std::deque implemented twice — ring-buffer variant (O(1) front/back, no reallocation, fixed capacity) and segmented-array variant matching the standard's reference-stability contract. The dual implementation surfaces where each approach's constant factors diverge under sequential versus random-access workloads.

• 2D ecosystem: 500+ bee agents across 2 hives, 9 respawning flower nodes; Boids flocking for swarm cohesion, foraging state machine for hive-to-flower routing
• Spatial hash grid reduced neighbour lookup from O(N²) to O(N) per frame — sustaining 60+ FPS at full agent count

Real-time supply chain risk platform built in 48 hours. Supply chains modelled as directed graphs; live risk propagation via Socket.IO with per-node LLM-backed assessments and geospatial rendering via MapLibre GL. Selected for 5-week Pre-Accelerator.

Autonomous robot navigating industrial pipe networks. A* path-planning with real-time obstacle avoidance for constrained pipe geometries. Functioning hardware prototype built and demonstrated to Microsoft & Deloitte judges in 48 hours. Did not place — awarded Most Pyrrhic for most ambitious attempt.