Astitva Jaiswal

Five minima, one descent.

• 01

  Multimodal evaluation & benchmarking

  Designing benchmarks for video and image models: rubric design with human review, failure-mode taxonomies, LLM-as-judge pipelines, inter-rater agreement and prompt optimisation. Probing where models fail, not just where they rank.

  RubricsLLM-as-judgeDSPy / GEPAVideoAgreement (κ)

  See CaptionBench
• 02

  Generative-media infrastructure

  Distributed job pipelines that serve many image, video and 3D generation models behind one API: queues and workers, multi-provider orchestration, usage and cost accounting, load testing and observability.

  FastAPIKafkaCeleryPostgresGCP / K8sBigQuery
• 03

  Agentic tooling

  Tool-grounded AI agents for engineering operations, with safety guards and human-in-the-loop pull requests. Every number in a reply comes from a tool call.

  Claude Agent SDKMCPClaude CodeSlack
• 04

  Code intelligence & RL environments

  Building verifiable software-engineering tasks for training and evaluating coding agents, using code dependency graphs, git-history mining and sandboxed verification.

  tree-sitterLSPDockerSWE-bench style
• 05

  Speech & language

  ASR, seq2seq translation and transliteration, Indic and code-mixed NLP, and distributed GPU training.

  PyTorchTransformersDeepSpeedNeMoWav2Vec2

Where models fail, not just where they rank.

• 2023A Critical Analysis of Approaches to Glaucoma DetectionIJARSCT · DOI 10.48175/IJARSCT-13871 ↗

• model↓ 19♥ 0

  Hindi → Hinglish code-switch

  ar5entum/marianMT_hin_eng_cs

  MarianMT fine-tuned to keep Hindi words in Devanagari and romanise English ones. BLEU 77.9 on a self-curated 100k set.

• model↓ 29♥ 0

  Bidirectional Devanagari ↔ Roman

  ar5entum/marianMT_bi_dev_rom_tl

  One MarianMT model that transliterates in both directions.

• model↓ 11♥ 0

  Roman → Devanagari

  ar5entum/bart_rom_dev_tl

  BART transliteration model, Roman script in, Devanagari out.

• model↓ 11♥ 1

  Devanagari → Roman

  ar5entum/bart_dev_rom_tl

  BART transliteration model, the reverse direction.

• model↓ 9♥ 0

  English → Hindi

  ar5entum/bart_eng_hin_mt

  BART machine translation.

• model↓ 17♥ 0

  Hindi → English

  ar5entum/bart_hin_eng_mt

  BART machine translation.

• dataset↓ 31♥ 0

  Hindi–English code-mixed corpus

  ar5entum/hindi-english-code-mixed

  100k verified rows pairing Devanagari Hindi with its code-mixed form.

• dataset↓ 69♥ 0

  Roman ↔ Devanagari transliteration corpus

  ar5entum/hindi-english-roman-devnagiri-transliteration-corpus

  Parallel transliteration pairs used to train the models above.

• 2023Semantic segmentation of plantsDetectron2 and YOLOv8 segmenting leaves, fruits and flowers on a hand-labelled dataset. ↗
• 2023Brain tumor detection on MRIPyTorch CNN on 12k scans, wrapped in an Electron + Flask app. ↗
• 2023Cancer data PCA + SVC97% accuracy classifier, with the PCA walk-through. ↗