Hemanth Kari

AI/ML Engineer, and Full-Stack and Frontend Engineer.

I work in Python, PyTorch, TypeScript, React, and Next.js. Lately a lot of LLMs and RAG.

Authorized to work in the US. No sponsorship. Open to relocation or remote.

Download ResumeGitHubEmail
hckariusa@gmail.com·+1 (312) 583-7830·huggingface.co/HemanthKari·ORCID

Experience

ASuiTech Solutions

Dec 2025 – Present · Remote

Full-Stack Software Engineer (Contract)

  • Rebuilt the production order-dispatch pipeline (200+ orders/day) as idempotent Kotlin Kafka stages with DynamoDB step state, designed for horizontal scalability. Failures self-heal via Kafka log replay instead of manual recovery scripts.
  • Diagnosed a recurring overselling bug in production on high-demand SKUs. The polling-based inventory sync had a race condition. Replaced it with an SQS / Lambda event flow plus Redis locks at the SKU level.
  • Moved the product and checkout pages to Next.js 14 server components. Core Web Vitals (LCP) on owned pages improved 12%. Separately, debouncing the React catalog filters and trimming the Elasticsearch query cut retrieval latency 30%.

Kotlin · Kafka · DynamoDB · Next.js 14 · TypeScript · SQS · Lambda · Redis · Elasticsearch

Horizon IT

Jun 2024 – Nov 2025 · Remote

Software Engineer

  • Migrated client state. Server state moved to React Query, UI state to Zustand. Cut redundant refetches by 40%. Added retry-with-backoff plus a unified error boundary so transient API failures stopped showing up as broken pages.
  • Accessibility pass on the assigned flows: Lighthouse score from 78 to 95 after replacing custom interactive widgets with semantic primitives and fixing keyboard / focus handling. Grew Jest and React Testing Library coverage by 22% so regressions caught in CI, not PR review.

React · TypeScript · Next.js · React Query · Zustand · Jest · React Testing Library

Deepcloud

Jan 2021 – Aug 2022 · Bangalore, India

Software Engineer

  • Wrote a TypeScript provisioning service that automated the full VM lifecycle for dev environments across the engineering org, replacing a brittle 30-minute manual checklist.
  • Hardened Python and Spring Boot microservices that had been compounding failures under load: circuit breakers on the failing-most calls, retry-with-backoff on the recoverable ones, graceful-degradation paths where partial data was acceptable. Profiled and debugged the PostgreSQL and MongoDB hot paths, added the missing indexes, rewrote a pair of N+1 query patterns on the worst route. Cut peak latency on the worst routes by 45%.
  • On-call rotation. Wrote runbooks, remediation scripts, and post-mortems for each recurring incident so the next page resolved with a known command. MTTR went from 30 to 10 minutes.

TypeScript · Python · Spring Boot · PostgreSQL · MongoDB

Projects

FluxionDock

Solo · 2025 – 2026 · In progress · US Provisional Patent in prep

Retrieval-augmented ML system for protein-ligand pose prediction

Retrieval-augmented flow-matching pose-prediction model in PyTorch (2.82 M research, 23.2 M shipping params). Retrieval keeps the generator small via analogous-complex conditioning; flow matching gives single-pass deterministic inference.

  • Research model: 1.60 Å mean Kabsch RMSD with 69.8% of poses below 2 Å on a 3,608-complex held-out validation set. Shipping baseline is evaluated leave-one-out, so retrieval cannot leak into the test conditioning. It reaches 1.49 Å / 75.3%.
  • Full training: 37 minutes over 10 epochs on a single A100.
  • Built the retrieval substrate: 72,199 RCSB PDB complexes, ESM-2 pocket embeddings plus RDKit Morgan ligand fingerprints, indexed in FAISS HNSW for sub-second lookup. CC-BY 4.0 on Hugging Face with reproducible GitHub Actions CI/CD, 900+ downloads.
  • Validated via SIFTS same-target evaluations: 65.1% same-UniProt at rank 1 (median cosine 0.982 vs 0.440 random).
  • Diagnosed why MoLFormer-XL (the preferred ligand encoder) would not load: its tokenizer files were incompatible with the pinned transformers version. Re-pinning the stack for one component was the wrong v0.1 trade. Fell back to RDKit Morgan, logged the active backend in config.json, kept the transformer path open for v0.2.

PyTorch · FAISS / HNSW · FastAPI · Next.js · ESM-2 · RDKit · A100

Dataset on Hugging Face DOI: 10.57967/hf/8408 Code (Zenodo)

Llama-3.1 Pro Coder v1

Solo · 2025 · 68.3% HumanEval · 1,200+ HF downloads

Code-specialist LLM fine-tune of Llama-3.1-8B-Instruct

Fine-tuned Meta Llama-3.1-8B-Instruct with QLoRA (rank 128) on 111,289 commercial-license code samples (48 M tokens). 4-bit quantization let the 8B model fit in 40 GB VRAM on one A100.

  • 68.3% pass@1 on HumanEval; 4-bit checkpoint released on Hugging Face, 1,200+ downloads.
  • Diagnosed a format mismatch between verbose CoT training data and HumanEval's zero-shot harness that was depressing scores even when underlying coding competence improved.

PyTorch · QLoRA · HF Transformers · 4-bit Quantization · A100

Model on Hugging Face DOI: 10.57967/hf/8407

Swale

Solo · 2024 – 2025 · Live on Google Play

Android launcher on Google Play

Android launcher shipped on Google Play with RevenueCat subscriptions and Firebase for analytics and crash reporting.

  • 100 downloads on Google Play.

Flutter · Dart · Firebase · RevenueCat · Android

Google Play

Publications

ORCID 0000-0001-5755-2242·Google Scholar

Kari, H., Bandi, S. M. S., Kumar, A., & Yella, V. R. (2023). “DeePromClass: Delineator for Eukaryotic Core Promoters Employing Deep Neural Networks.” IEEE/ACM Transactions on Computational Biology and Bioinformatics, 20(1):802–807.

First-author CNN-LSTM, 93.6% test accuracy on C. elegans. 11 Google Scholar citations (8 received in 2025); cited in Trends in Biotechnology (Cell Press, 2024 IF 14.3).

doi:10.1109/TCBB.2022.3163418

Kari, H. (2026). “Pocket and ligand embeddings for 72,199 protein-ligand complexes from the Protein Data Bank.” Sole-author Data Descriptor under bioRxiv screening.

Pipeline source on Zenodo.

doi:10.5281/zenodo.19954940

Datasets and Models

[Dataset, 2026]rcsb-ligand-retrieval-v0.1: 72,199 protein-ligand complex pocket and ligand embeddings. Hugging Face Datasets · CC-BY 4.0 · 900+ downloads.

doi:10.57967/hf/8408

[Model, 2025]Llama-3.1-Pro-Coder-v1: QLoRA fine-tune of Llama-3.1-8B-Instruct (68.3% HumanEval). Hugging Face Models · 1,200+ downloads.

doi:10.57967/hf/8407

Skills

Languages
Python, Java, Kotlin, TypeScript, JavaScript, SQL
Backend
Node.js, Spring Boot, FastAPI, Express.js, REST, GraphQL, Kafka, Redis, SQS, Lambda
Frontend / Mobile
React, Next.js, Tailwind CSS, Jest, React Testing Library, Flutter, Firebase
Infra / DevOps
AWS (EC2, Lambda, S3, DynamoDB), Docker, GitHub Actions, PostgreSQL, MongoDB, Elasticsearch, Observability
ML / AI
PyTorch, HF Transformers, LoRA / QLoRA, FAISS / HNSW, RAG, Embeddings, RDKit, ESM-2, HumanEval, 4/8-bit Quantization

Education

M.S. Computer Science

Aug 2022 – May 2024

Purdue University Northwest, Hammond, IN · GPA 3.5 / 4.0

Coursework: machine learning, distributed systems, algorithm design, database systems, computer security.

B.Tech. Biotechnology

Jul 2017 – May 2021

KL University, India

Honors

Hack The Box Top 100 Hall of Fame

Verified badge · December 2021

Global cybersecurity leaderboard.

First-author IEEE/ACM TCBB

Cited in Trends in Biotechnology (Cell Press, IF 14.3)

DeePromClass, 2023. 11 Google Scholar citations, 8 of them in 2025.