Not what a model outputs — how the system decides, executes, and holds under load.

Ashwin Gupta

AI Engineer
SkanAI
Jun 2026

MLOps & GenAI · IIIT Bangalore · PyTorch · LLMs · RAG · GCP

↓ Scroll

Inference is easy. Everything around it isn't.

Honest where it matters. Available when it's hard.

Mechanical engineering by training — which meant learning to ask why a system fails before asking how to build it. AI hit in second year like a realisation, not a subject: software that understood language was a new class of thing.

Production changed the picture fast. The GenAI core — prompts, basic RAG, API calls — is learnable in three months. Everyone builds it. The real gap is what surrounds the model: the routing logic, the concurrency architecture, the observability that tells you what actually broke and when.

At Coforge on the HSBC Conversational Analytics project: GIL'd threading replaced with CPU-pinned parallel instances, asyncio + uvloop across the full pipeline. 7× session capacity. $1.3M annualised savings. MTTR from ~1hr to ~10mins.

Inference as a System

Most teams ship inference as a function call. The real questions — p95 latency, 10× load, what happens when a backend goes down — are architecture questions.

Execution Under Constraints

Systems that perform in demos often don't survive production. Real constraints — latency budgets, VRAM ceilings, cost per token — are known at design time.

Physics-Informed Scientific ML

Data-driven physics models aren't data problems — they're structure problems. Embedding PDEs into the objective is what makes sparse data sufficient.

What I don't do

  • I don't ship AI wrappers dressed as products.
  • I don't build for its own sake. The system has to earn what it costs to run.
  • I don't take off-the-shelf work. If the implementation is a Google search away, I'm not the right person.

The trajectory.

AI Engineer SkanAI Jun 2026 – Present
  • Applied LLMs & agentic systems
  • Process intelligence platform
AI Engineer Coforge Jun 2024 – Jun 2026
  • Conversational Analytics (HSBC) — SBC→STT→LLM on GCP/RHEL
  • GIL fix: CPU-pinned procs + asyncio/uvloop · 20→140–160 sessions/VM
  • Compute: $118K→$8K/month (~$1.3M/yr) · MTTR 1–2hr→~10min
  • Best Team Award · Pat on Back · Keep It Up Award
Data Scientist Gida Technologies Jan 2023 – May 2024
  • Here.app (HDFC ERGO) — 163-lang multilingual RAG · 97% factual accuracy
  • Prismforce Skill Graph — +30% relevance · sub-50ms on NVIDIA T4
Head of Machine Learning IISc — NMCAD Lab Jan 2022 – Sep 2022
  • eVTOL aerodynamic & structural optimisation under Prof. Harursampath
  • Physics-constrained surrogate ML · 5 projects across fluid, structural & thermal
B.E. Mechanical Engineering BMS College of Engineering Aug 2019 – May 2023
  • Founder & mentor — Augment.AI · Best Outgoing Project '23 (PINNs)
  • IEEE Joint Secretary · chapter ranked #2 globally
Executive Diploma, AI & ML IIIT Bangalore Oct 2025 – Mar 2027
  • Dual specialisation — MLOps, GenAI & Agentic AI

Proof, not promises.

Session capacity increase
$1.3M
Annualised cost savings
~10 min
MTTR (from ~1–2 hr)
97%
Factual accuracy (RAG)
sub-50ms
Inference on NVIDIA T4
+30%
Recommendation relevance

Systems that had to hold.

ashwingupta.dev — Design Handoff to Production

Shipped Personal
Problem Original portfolio claimed performance engineering while shipping 400 animated DOM nodes and a 2 MB JPEG hero.
System Rebuilt as a three-layer spatial interface — all visual effects collapsed into a single Canvas RAF loop, offscreen pre-rendering, lazy loading, WebP.
90% image reduction · 72% JS cut · 400 DOM nodes eliminated · frame time 18–25ms → 4–6ms

PageIndexOllama — Local-First Fork of PageIndex

Shipped Open Source
Problem Tree-RAG was hardwired to one provider contract — completion differences silently corrupted recursive traversal.
System Provider-routing layer with finish-reason normalization; prompt externalization, bounded concurrency, hierarchical fallback.
Fully offline tree-RAG across Ollama, llama.cpp, and vLLM — no external API keys required.

Azure Infrastructure Documentation Engine

Client Delivery Coforge
Problem Azure docs relied on manual exports and hand-drawn diagrams — every project took 2–3 days and drifted from live state.
System Live-state extraction pipeline — subscription scan, topology mapping, security config analysis auto-generate SDDs and PlantUML.
2–3 days → ~2–3 hours · 104 resource groups per engagement · zero fabricated components

Graph-Based Skill Recommendation Engine

Client Delivery Prismforce
Problem Skill recommendations ignored hierarchical relationships; taxonomy changes forced full batch retraining; missed sub-50ms SLA.
System Weighted directed graph over multilevel skill hierarchies with typed edges and lightweight scoring.
+30% relevance · sub-50ms inference on one NVIDIA T4 · live taxonomy updates without retraining

In their words.

"Ashwin showed a great flexibility and stretched to complete a challenging task which resulted in client's delight. He was able to code a logic which client's other partner could not do it."

Arun Kumar Vastrakar
Senior Delivery Director · Coforge · Nov 2024

"Ashwin's work on the telephony ingestion layer — PJSIP-based, highly stable, low-latency SIP call-handling at scale — was technically precise. Highly recommended for roles in VoIP engineering or conversational-AI infrastructure."

Raja Sekhar Amirapu
Senior Technical Architect · Coforge · Nov 2025

"His expertise in scalability, testing, and debugging proved invaluable. He played a crucial role in designing and implementing the scalability logic, ensuring our infrastructure could handle increasing workloads efficiently."

Snehasish Chakraborty
GCP Infrastructure Engineer · HSBC · Feb 2025

"Ashwin is a risk-taker, never shying away from trying innovative approaches — and what sets him apart is his ability to convert those risks into successful implementations."

Tulsi Patro
AI Engineer · Gida Technologies · Nov 2023

What I run in production.

Profiled under load. Not just imported.

Backend & Systems

Distributed systems Microservices REST / OpenAPI Kafka / Pub/Sub Concurrency & perf Redis Grafana / Prometheus Fault tolerance

Real-Time & Voice

PJSIP / PJSUA2 Kamailio SIPp TLS · DTLS/SRTP

Data & ML

FastAPI PyTorch LLMs / RAG FAISS / ANN NetworkX Chroma / HNSW Spark

Profiling & Infra

Scalene line_profiler Memray GCP / Azure Packer / GCE Locust