I'm Rushabh

Architected a tree-shakeable TypeScript WebRTC SDK end-to-end as sole SDK engineer. Cut enterprise integration from 4–5 days to under 2 hours and delivered a 50-room race-condition-free state layer.

• ▸ Architected a tree-shakeable TypeScript WebRTC SDK with sub-200ms p50 latency; a plugin interface enabled 3 enterprise integrations without core SDK modifications.
• ▸ Eliminated state desync across 50 concurrent rooms by replacing last-write-wins Firestore with a server-timestamp reconciliation protocol, validated with a race-condition harness.
• ▸ Cut client integration from 4–5 days to under 2 hours as sole SDK engineer; aligned product and engineering on transport tradeoffs and delivered no-code onboarding end-to-end.

Launched a multi-tenant voice-agent platform serving 4 enterprises at 2.3K concurrent calls/day with 99.9% uptime, and redesigned the WebSocket turn-taking pipeline to cut TTFB to 800ms.

• ▸ Launched multi-tenant voice platform serving 4 enterprises at 2.3K concurrent calls/day with 99.9% uptime, automating 70% of manual SDR outreach workflows.
• ▸ Cut TTFB to 800ms by redesigning the WebSocket turn-taking pipeline with a custom VAD; deployed stateless GCP Cloud Run microservices with zero-downtime auto-scaling.

Co-architected a CAD-drawing validation platform under a semester-long engagement with ISGEC Heavy Engineering. Two-phase pipeline: synchronous rule engine + async LLM enhancement pass.

• ▸ Designed a two-phase pipeline: Phase 1 (parsing + 8-rule engine) completes in ~1.8s on a real 7.4MB drawing (~17.8K entities); Phase 2 runs as an async LLM enhancement pass with a hard timeout.
• ▸ Co-built an 8-check geometric validation engine using Shapely spatial indexing (STRtree) for sub-quadratic clash detection, optimal assignment for 1:1 entity matching, and RapidFuzz for fuzzy ID reconciliation — covered by 49 pytest cases across the rule engine, parser, revision ledger and end-to-end flow, and tested against real PostgreSQL rather than an in-memory substitute.

Industry engagement with Bekaert: the hardware and controls design for a 90 kg four-wheel line-following AGV — motor, driver and battery sizing against a single mass budget, encoder-closed-loop wheel-speed sync on STM32, and the telemetry dashboard the platform is operated from.

• ▸ Sized the complete drive chain against one mass budget: 90 kg target closed to a 97–98 kg design mass with a 25 kg chassis, four 21 Nm motors on independent drivers at 36 A total rated current, and a 2S4P LiFePO4 pack (25.6 V, 120 Ah, 3.07 kWh) for a 6–8 hour duty cycle.
• ▸ Closed the wheel-speed loop in firmware on STM32 using incremental A/B magnetic encoders on all four motors, so the commanded value is a speed rather than a PWM duty cycle — without it the left and right wheels drift apart and the rover curves off the line within seconds.
• ▸ Built the operator telemetry dashboard the platform runs on: per-motor driver, PWM, direction and encoder counts, LiDAR stop/slow/clear safety zones, a per-driver bench test, and session recording — with a documented open constraint on the driver voltage ceiling that still has to be closed before the pack is built.

Industry-engagement project for Yantrikisoft: a multi-tenant retail analytics platform on a 73,100-row dataset — a 45-feature gradient-boosting demand model with conformal intervals, PSI drift monitoring, a three-estimator causal ensemble that reports its own disagreement, and an inventory simulator with an RL policy measured against a heuristic baseline.

• ▸ Built the demand model end to end on 73,100 rows across five stores and twenty products: 45 lag, rolling, calendar, price and competitor features into a gradient-boosted regressor, reaching MAE 71.6 and RMSE 92.0 on a held-out window against a naive lag-1 baseline MAE of 89.2 — with zero-demand rows treated as censored rather than as zeros.
• ▸ Triangulated promotion effects with three independent estimators (propensity matching, doubly-robust residual regression and difference-in-differences) combined by inverse-variance weighting, and surfaced the disagreement rather than hiding it — the three methods spread 187% on the shipped data, which the platform reports as a warning.
• ▸ Replaced fixed reorder thresholds with a SimPy inventory simulator and a PPO policy trained via Stable-Baselines3, measured against the (s,S) baseline on the same episodes: 12.8% lower total inventory cost at an equal service level, plus PSI drift monitoring over six features to catch the model going stale.