Hi, I am
Software Engineer at RedGraphs and AI researcher based in Delhi. I build NLP and backend systems and work on AI-assisted mathematics, with a focus on traceable evidence, reproducible results, and formal verification. Previously a Research Fellow at Lossfunk, with research experience in ECG classification, parameter-efficient fine-tuning, and molecular simulations.
Automated verification and scoring for five problems in mathematics, computer science, and computational biology. Built constraint validation, scoring, and percentile rankings for the conference's Verifiable Problems Track.
Supervised fine-tuning and GRPO with LoRA on Qwen using MedMCQA. Implemented rewards for answer correctness and structured outputs, and evaluated baseline, SFT, and GRPO checkpoints.
How many regions can overlapping A-shaped figures create? Established the exact formula for constrained long-legged A arrangements, with a visual proof, rational certificates through n=16, and reproducible verification.
Co-authored a two-graph approach to Neil Sloane's lollipop problem with Paras Chopra. Closed Paulsen's remaining gaps through n=17 and determined the exact value aL(19)=1076.
Build NLP and backend pipelines that extract company relationships from regulatory filings, financial disclosures, and public information. Develop entity resolution and evaluation workflows for precision, recall, and evidence traceability, and work on backend performance, observability, and production reliability with AWS and PostgreSQL.
Worked on AI-assisted mathematics across combinatorics, integer sequences, and formal verification. Co-authored research refining bounds for Neil Sloane's lollipop problem, developed exact results and computational certificates for OEIS sequences, and formalized results in Lean 4 and Mathlib with an emphasis on reproducibility and verifiability.
Helped organize the Conference for AI Scientists, focused on AI-assisted scientific discovery. Contributed to the Verifiable Problems Track and the infrastructure for evaluating submissions in mathematics, computer science, and computational biology.
Worked on NLP pipelines and conversational AI for a voice-first mental health support application, using RAG, ChromaDB, and Docker with a focus on experimentation and evaluation.
Worked with Prof. Anubha Gupta on parameter-efficient fine-tuning (PEFT) and LoRA extensions for custom Vision Transformers trained on over 10 million ECGs.
Developed ECG-based subclass classification for cardiovascular diseases using signal processing. Conducted a review of risk calculators for Multiple Myeloma staging.
Under Prof. Prabal Maiti, researched binding affinities in antibody-antigen binding using ChimeraX, Modeller, and GROMACS for MD simulations.
Built a ML model to classify educational YouTube videos. Handled pipeline: scraping (YouTube API), cleaning (BeautifulSoup), and modeling (ML.NET/C#).
Proof pages and a compact ledger of published contributions, credited work, and drafts under review.
Visual explanations, manuscripts, exact certificates, and reproducible verification.
The complete all-n proof for A397182, with both 55-region figures, exact rational certificates through n=16, and standard-library verification scripts.
A visual account of the independent finite-certificate proof that a(9)=20; the approved OEIS edit recovered Harborth's 1985 source for a(8)=16.
Exact values through n=25, with connected witnesses and matching upper proofs, proposed as OEIS A398723.
Approved contributions and research credited in OEIS records or their revision histories.
With Paras Chopra, closed Paulsen's remaining gaps through n=17 and at n=19, using a two-graph refinement of the crossing obstruction.
Proved two exact values, tightened the open n=16 interval, and added a private-chain upper bound that works in every dimension.
Identified the statistic, proved three exact identities, and supplied the approved b-file through n=1000.
Proved the exact value of a(3) by exhaustive minimality checking and added a constructive upper bound for a(4).
Corrected two carried terms from 301, 352 to a(24)=302 and a(26)=353, against Chu–Stuckey's published tables.
Added an independent double-counting check of Mutoh's value a(5)=905,697,107,804,160, using the count for the complement of a fixed cycle.
Recasts the lollipop crossing problem through two interacting extremal graphs, closing Paulsen's remaining gaps through n=17 and determining aL(19)=1076.
Conducted predictive modeling for anti-epileptic drug outcomes using patient data. Utilized six Machine Learning algorithms, achieving over 70% accuracy in predicting drug responses in the North Indian population.
Focus: IT & Mathematical Innovations, Computational Biology.
