I build production AI systems: multi-agent platforms, retrieval pipelines over 360K+ documents, and inference stacks that run 9x faster. First inventor on a published patent for multimodal healthcare AI.
I am Aditya Jethani, a Machine Learning Engineer based in India, working on agentic AI, retrieval-augmented generation (RAG), LLM fine-tuning, and the ML infrastructure that takes models to production. I measure everything and optimize until latency, accuracy, and cost targets are met.
- Patent: First inventor on a published Indian patent, AI-Powered Multimodal Healthcare System for Elderly Care, whose core novelty is a unified multimodal vector store fusing text, voice, and image into a single embedding space.
- Now: Building governed enterprise AI platforms and multi-agent systems that query live business systems, plus applied research on deterministic document parsing.
- Open source: Contributor to Mem0, OpenLIT, OpenAI Agents SDK, Outlines, and Nous Research's Hermes Agent.
- Competitive programming: LeetCode Knight, rating 2071 (top 1.75%), AIR 61 in Weekly Contest 462 among 35,000+ participants.
- Education: B.Tech Computer Engineering, CGPA 9.44/10 · GATE CS 2026: 95.61 percentile.
| Work | Details | Status |
|---|---|---|
| AI-Powered Multimodal Healthcare System for Elderly Care | Indian Patent App. No. 202521086225 · first inventor (mentor Dr. Debabrata Swain, PDEU). Novelty: a unified multimodal vector store fusing text, voice, and image into one embedding space via an attention-based feature-alignment layer, with a Vision Transformer for medical imaging and RAG grounded in user history. | Patent Published |
| Advanced Forecasting of Solar Power Generation using Bi-LSTM with Fourier Features | First author. Bi-LSTM with 25 Fourier term-pairs for multi-horizon PV forecasting across two Indian plants (R² = 0.861, RMSE = 128.09 kW), plus an adversarial degradation model for maintenance recommendations. | Under review · IEEE Transactions |
| Efficient Labelling of Hinglish Datasets for Hate-Speech Classification | Ensemble-consensus labelling (Twitter-RoBERTa, BETO/mBERT, VADER) with iterative refinement for low-resource code-mixed Hinglish, reaching 86% accuracy. | Academic research |
| Deterministic Document Parsing that Outperforms VLM Parsers | A zero-ML, failure-mode-driven PDF parser reaching CER 0.0135 and token-F1 0.99 on native PDFs. | In preparation |
| What | Result |
|---|---|
| LLM inference pipeline | 9x faster (110s to 12.2s, 139.6 tokens/sec) |
| Agentic RAG platform | 360K+ multilingual documents, 4.3x throughput |
| In-house PDF parser | Beats PyMuPDF on 23 of 34 metrics (token-F1 0.94) |
| Enterprise multi-agent assistant | 7 production agents over live ERP, accounting, CRM and regulatory data |
| Client agent delivery | 70% faster |
| Natural-language-to-SQL | Governed SQL + charts in under 2s |
| Clinical tongue-analysis CV | 94.2% color, 91.7% coating accuracy |
Real fixes and features in production AI infrastructure, not tutorial clones.
| Project | What I contributed |
|---|---|
| mem0ai/mem0 | Fixed the Weaviate vector store writing OutputData model fields as properties on update, corrupting stored memories. Merged. |
| dottxt-ai/outlines | Authored documentation for the LM Studio model backend. Merged. |
| openlit/openlit | Groq audio instrumentation for the Python SDK (1,024 lines, 164-case test suite, OpenTelemetry semantic conventions), plus reasoning-content capture for Groq and Ollama. |
| openai/openai-agents-python | Fixed metadata leaks in AdvancedSQLiteSession.clear_session and pop_item. |
| NousResearch/hermes-agent | Fixed the SQLite session store leaking on wake-gate skip paths in the cron loop. |
| wasmerio/Python-Scripts | Profanity-checker utility. Merged. |
Plus 15+ merged PRs across Hacktoberfest and GSSoC.
Setu: Kumbh Lost & Found Network · Next.js Flutter FastAPI Sarvam AI Docker
Offline-first, voice-first platform reuniting families separated at the Kumbh Mela. One-tap speech intake in 23 Indian languages. Matching engine: Recall@1 87%, Recall@5 97% with hashed-PII privacy design.
Company Brain · FastAPI pgvector Redis Docker
Multi-tenant RAG platform with hybrid semantic search and citation-grounded answers, served through async microservices behind a CI pipeline.
TalentScout AI · LangChain LangGraph Transformers
Autonomous screening agent with dynamic question generation and running confidence scoring. 85% validation accuracy, 60% less evaluation time.
Prompt Detective · OpenCV CLIP FastAPI
Reverse-engineers the prompts behind AI-generated media, with async job processing and idempotent uploads.
Languages
Agentic AI & LLMs
ML & Deep Learning
Infrastructure & Observability
- LeetCode Knight — rating 2071 (top 1.75%), AIR 61 in Weekly Contest 462 among 35,000+ · CodeChef rank 936
- GATE CS 2026 — 95.61 percentile
- Google Solution Challenge 2025 — Top 105 globally (4,000+ teams)
- ByteVerse 2025 (IIT Patna) — Top 10 · JPMorgan Code for Good 2024 — Finalist · Smart India Hackathon 2023 — Finalist
- Anthropic certified — Claude, MCP, Agent Skills · NVIDIA Deep Learning Institute
- Chairperson & Advisor, ACM-PDEU Student Chapter · Tech Lead, GFG-PDEU