Ashwin Gupta

GIL-bound threading serialised all audio onto one core on a 32-core VM - 20 concurrent calls, 31 cores idle, CPU utilisation capped below 20%.

CPU-pinned parallel processes via Linux taskset escaped the GIL; asyncio + uvloop replaced threading - each SIP session became a coroutine across SBC, STT, and LLM stages.

Packer automated GCE image builds; SIPp validated 2,000 concurrent users; n2-standard-32 → c4-standard-8 gave 30–40% longer transcripts; GCP log correlator scanned 250K+ lines in <5s, cutting MTTR to ~10 min.

20 → 140–160 sessions/VM (7×); 1,600+ peak at <300ms E2E, <5% packet loss; compute $118K → $8K/month (~$1.3M/yr saved); Best Team Award - HSBC Account.

Local inference routing is stateless - prior outcomes ignored, blind dispatch repeated indefinitely; no unified API surface across modalities.

19 backends across 7 modalities - text, STT, TTS, image gen, embeddings, vision, reasoning - under one OpenAI-compatible API; not a proxy, a control plane.

Every request classified, scored across 6 dimensions (capability, performance, resource, load, reliability, context), queued in Redis, dispatched, observed. EWMA weights per (backend, task_type, complexity) persist across restarts.

Routing accuracy compounds with deployment time - versioned policy, reward signals, and ReplayEngine regression gating. No retuning.

Standard chatbots failed on structured vehicle spec data - inconsistent answers drove manual escalation at HDFC ERGO, one of India's largest general insurers.

Multilingual RAG pipeline with structured retrieval, language-aware chunking, and per-language factual validation - 163 languages as first-class targets, not as post-processing.

QA-gated retrieval validates lookup quality before generation; structured vehicle database with image-linked attributes as single source of truth across 163 languages.

~97% factual accuracy on vehicle spec queries; autonomous sales-support resolution at HDFC ERGO scale; reduced manual escalation across full product catalogue.

Purely data-driven physics simulation demanded large labeled datasets - sparse data let models ignore governing equations, producing physically implausible solutions.

Dual-loss PINN framework embedding PDE/ODE constraints directly into the optimization objective beside data loss - physics constraints act as the regularizer.

Validated across 6 benchmarks: Burgers' equation, 1D heat conduction, fixed-fixed column deflection, cantilever deflection, transient cooling under Neumann and Dirichlet conditions.

Stable convergence across fluid, structural, and thermal domains with limited labeled data; Best Outgoing Project - Mechanical Engineering, BMSCE, 2022–23.

Physics-informed models assume the governing equation is known - the harder upstream problem is regime classification: which regime applies to the observed field.

Dual encoders map symbolic descriptions (PDE) and field observations into a shared latent space; regime recognition emerges from cross-modal alignment, not hand-coded rules.

Contrastive objective pulls matched (description, observation) pairs together, pushes mismatched apart - PHYSCLIP as a perception layer upstream of PINN-style enforcement.

Latent proximity carries physical meaning - nearby embeddings indicate similar regimes, enabling interpretable identification under partial observability.

MCQ distractors must be semantically plausible - factually wrong alone fails to separate understanding from guessing

BERT surfaces salient spans as pivots; proper nouns anchor pivots; WordNet and ConceptNet generate nearby alternatives through sense traversal

WordNet supplies hypernym→hyponym chains with sense disambiguation; ConceptNet adds part-of and related-to structure when coverage thins; fallback ensures distractor coverage

Semantic distance governs selection - distractors must occupy the same conceptual neighbourhood without matching the tested senses; proximity without identity is the key constraint

Research copilots generate fluent text without evidence traceability - grounded synthesis and hallucination look identical; no claim can be audited back.

Five locked MCP services cover literature mapping, contradiction detection, hypothesis critique, evidence extraction, and assembly via schema-defined interfaces.

Only hypothesis critique is agentic - bounded to five iterations; all other stages are deterministic with provenance tracked through typed artifacts.

Each claim is bound to source evidence; contradiction detection marks where consensus breaks, keeping outputs falsifiable and useful beyond sessions.

The original portfolio claimed performance engineering while shipping 400 animated DOM nodes and a 2 MB JPEG hero - self-defeating on load.

Rebuilt as a three-layer spatial interface - all visual effects collapsed into a single Canvas RAF loop, offscreen pre-rendering, lazy loading, WebP preloads. Structural optimization, not cosmetic.

Extended with an ambient HUD system - geolocation-to-nearest-airport clocks, scroll-depth exploration tracker with color-staged progress arc, normalized mouse XY - all persistent across Astro View Transitions.

90% image reduction • 72% JS cut • 400 DOM nodes eliminated • frame time 18–25ms → 4–6ms • 17 pages tracked at scroll-depth resolution.

Tree-RAG was hardwired to one provider contract - completion differences silently corrupted recursive traversal; failures surface only at collapse.

Added a provider-routing layer with finish-reason normalization, so traversal depends on stable internal contracts, not whichever runtime answered.

Prompt externalization, bounded concurrency, and hierarchical fallback stabilize long-document runs on local models with uneven outputs and limited memory.

Fully offline tree-RAG across Ollama, llama.cpp, and vLLM - provider switching is transparent, with no external API keys required.

Every incoming clickstream needed a new dashboard; process artefacts were scattered across local folders, and Markdown reviews required another tool.

Built a reusable process workspace that brings clickstreams, case documents, frame metadata, and phase artefacts into one interactive dashboard.

One command onboards a process without copying source data. All eight phases are live, keeping the full artefact lifecycle in the same workspace.

Study and artefact kickoff: 3–4 hours → 10–15 minutes. Adopted as the standard dashboard across the team.

Azure docs relied on manual exports and hand-drawn diagrams - every project took 2–3 days and drifted from live state.

Built a live-state extraction pipeline - subscription scan, topology mapping, and security config analysis auto-generate SDDs and PlantUML from live resource evidence.

Few-shot prompting grounds generation in extracted inventory; guardrails reject any component without a matching live resource - fabrication blocked from governance docs.

2–3 days → ~2–3 hours • 104 resource groups per engagement • zero fabricated components • manual PlantUML authoring removed.

Skill recommendations ignored hierarchical relationships, taxonomy changes forced full batch retraining, and live inference missed the sub-50ms SLA.

Built a weighted directed graph over multilevel skill hierarchies with typed edges, lightweight scoring - structure, not retraining, drives relevance.

Dynamic node insertion and deterministic traversal keep the graph current; latency was profiled at the 99th percentile under production load.

+30% relevance • sub-50ms inference on one NVIDIA T4 • taxonomy expansion no longer required batch retraining • live updates stayed current.