Abhinav Gupta

Software Engineering·Research·Sidequests

M.TECH COMPUTER SCIENCE · IIT DELHI · NEW DELHI

Abhinav Gupta

I work in the space between building systems and proving they work — query engines, evaluation design, and the parts of machine learning that have to be measured rather than asserted.

My M.Tech thesis at IIT Delhi measures how closely audio language models reproduce clinician-assigned scores on standardised aphasia assessments. Before graduate school I spent close to four years building distributed data systems — three years at Fractal.ai on portfolio optimisation and performance analytics, then at American Express on the global prospect data platform.

Currently

Running the audio-English arm of the benchmark — comparing Whisper and Deepgram transcripts as scoring evidence across the four assessment tasks. September 2026

Abhinav Gupta

01Research

MAY 2026 — PRESENT

M.Tech thesis
M3RG, IIT Delhi

Supervised by
Rahul Garg,
N M Anoop Krishnan

Python · PyTorch
Whisper · Deepgram
scikit-learn · SHAP

Aphasia‑LLM

Benchmarking audio language models against clinical aphasia assessment

Aphasia is an acquired language disorder, most often following stroke, and its severity is graded by clinicians administering standardised assessments. This work measures how closely current audio and text language models reproduce those scores from the same recordings.

The benchmark spans four AphasiaBank instruments over cohorts of 110 to 130 participants, crossing seven scoring models with two ASR systems and five input conditions. Transcripts are frozen once per participant and reused unchanged across every condition, so ASR effects and scorer effects can be separated rather than confounded. Reference scores are read at evaluation only.

4

AphasiaBank
instruments

7

scoring
models

2

ASR
systems

5

input
conditions

02Experience

DEC 2024 — JUN 2025

Gurgaon

Python · SQL
PySpark · Hadoop
GCP

American Express

Software Engineer — Global Data Platform & Prospect Intelligence

  • Built ingestion pipelines for the global prospect data platform, orchestrating workflows over 5.6M+ records.
  • Implemented fuzzy-matching to reconcile multi-source vendor records against internal CRM databases.
  • Added automated deduplication, raising downstream batch throughput by roughly 40%.

AUG 2021 — DEC 2024

Bengaluru

Python · PySpark
Azure · Databricks

Fractal.ai

Software Engineer — Portfolio Optimisation & Performance Analytics

  • Built distributed PySpark systems over 10M+ records; custom partitioning cut job latency by 75%.
  • Built a SKU recommendation engine combining KNN clustering with MINLP optimisation for a Fortune 500 client.
  • Architected ETL batch pipelines and automated portfolio refresh, removing a standing manual step.
  • Received the internal Star Award in 2022 and 2023.

03Projects

JAN — APR 2026

IIT Delhi
Kaustubh Beedkar

Rust · TPC-H

Source ↗

Memory-constrained query engine

  • Out-of-core TPC‑H engine under a 64 MB ceiling, backed by an O(1) LRU buffer pool.
  • Volcano-style iterators with projection pushdown and hybrid joins over pipe-based I/O.
  • Executes multi-gigabyte workloads with no memory spills across nested operators.

DEC 2025 — MAR 2026

IIT Delhi
Kolin Paul

C++ · Flex
Bison · Linux

Source ↗

Shell, compiler and virtual machine

  • POSIX-compliant shell in C++ with multi-stage pipelines, process isolation, signal handling and job control.
  • Flex/Bison compiler lowering a C-like language to an AST and then to a custom bytecode.
  • 64‑bit stack VM with mark-and-sweep garbage collection.

SEP — NOV 2025

IIT Delhi
Mausam

C++

Source ↗

Adversarial game agent

Rivers & Stones

  • Iterative-deepening minimax with α–β pruning under fixed per-move time budgets.
  • Zobrist hashing, transposition tables and move-ordering heuristics to extend search depth.
  • Evaluation combining positional mobility with Manhattan-distance territory control.

FEB — APR 2025

Scaler

Python · FastAPI
PostgreSQL · Redis
Kafka

Distributed e-commerce platform

  • Event-driven services for order, inventory and payment.
  • Saga pattern for distributed transactions, with idempotent consumers and automated retries.
  • Held p95 latency under 50 ms at ~2,000 req/s via Redis caching, multi-column indexing and async event loops.

2019 — 2021

IIT Guwahati

Sidequests

Undergraduate detours that taught me the most

  • Linux process debugger — breakpoints, register inspection and single-stepping over ptrace.
  • AR tag detection — contour detection, homography and perspective warping from first principles.
  • Contactless fingerprint recognition — image enhancement and minutiae matching.
  • Protein toxicity prediction — pipeline over ~42K UniProtKB proteins; ANN at ~93% validation accuracy.

04Writing

From objects to services

SEP 2026

The SOLID principles stated plainly, then the harder question — which of them still mean anything once the method call becomes a network call.

Freezing the evidence

SEP 2026

Separating transcription error from judgement error when a language model scores a clinical assessment from audio — and what an artifact has to be for the separation to hold.

05Education

IIT Delhi — M.Tech, Computer Science & Engineering

2025 — 2027

CGPA 8.33 · department rank 5

IIT Guwahati — B.Tech, Chemical Science & Technology

2017 — 2021

CGPA 8.48 · Samsung Fellowship 2021, top 10% of bachelor theses

Teaching assistant — COL106 Data Structures & Algorithms, COL7001 System Concepts

IIT DELHI

Lab tutorials, assignment design and grading.

06Recognition

Department rank 5, M.Tech Computer Science & Engineering

IIT DELHI

Samsung Fellowship — top 10% of bachelor theses

2021

Star Award, Fractal.ai — performance and client appreciation

2022, 2023

JEE Advanced — top 3% of 2.1 lakh candidates

2017

07Technical

Languages
Python, C, C++, Rust, SQL, Java
Systems
Linux, PostgreSQL, Redis, Kafka, FastAPI, PySpark, Databricks, Azure
Machine learning
PyTorch, scikit-learn, NumPy, Pandas, SHAP
Foundations
Data structures & algorithms, operating systems, databases, system design

08Contact

I am open to research collaborations and to software and machine learning internships. Email is the most reliable way to reach me.

The first principle is that you must not fool yourself — and you are the easiest person to fool.

Richard Feynman