I'm Shreyan Gupta. I design infra systems for AI agents: memory, multi-agent orchestration, and the UX people actually touch. Recently shipped hermes and clawdown. Previously built RAG + compliance AI at Exelon.
Based in Philadelphia. Graduating June 2026. Open to full-time roles.
░░░░░░░░░░░░░░░░░░░░
00:00tap to see color
I care about the things most people skip. The edge cases, the impossible deadlines, the parts of a codebase nobody wants to touch. I'm currently shipping my way through building things that matter.
RESEARCH
An LLM-Based ETL Architecture for Semantic Normalization of Unstructured Data
Published in
MIT×ばつIEEEURTC 2025
An LLM-Based ETL Architecture for Semantic Normalization of Unstructured Data
Shreyan Guptasole author
scroll
EXPERIENCE
Drexel
DREXEL UNIVERSITYB.S. Computer Science · 2022 — 2026
DREXELB.S. Computer Science2022 — 2026
Drexel
DREXEL UNIVERSITY@ Undergraduate Researcher
PHILADELPHIA, PA
DEC 2025 — PRESENT
PART-TIME
DEC 2025 — PRESENT
PART-TIME
2 research roles — TeX-Base Lab, VI Lab
click to expand ↓UNSW
UNSW · GLOBAL EXCHANGEsydney, australia · sep — dec '25
HANDSHAKE@ MOVE AI Fellow
SEP 2025 — PRESENT
CONTRACT
SEP 2025 — PRESENT
CONTRACT
evaluating frontier LLMs on reasoning + accuracy + safety & writing structured r...
click to expand ↓EXELON@ Data Scientist
PHILADELPHIA, PA
SEP 2024 — MAR 2025
CO-OP
SEP 2024 — MAR 2025
CO-OP
- ▪automated SPCC risk assessment across 12.5k assets. LLM structuring hit 99.4% accuracy, saved ~1ドル.2M/yr in regulatory overhead
- ▪built PySpark pipelines processing 30M+geospatial records with H3 clustering on Azure Synapse for predictive modeling
- ▪designed RAG ingestion for 3.5k docs at ~35ms per file, shipped as Dockerized services on Azure with Terraform + CI/CD
ELAR INDUSTRIES@ Software Developer
MAR 2023 — SEP 2023
CO-OP
MAR 2023 — SEP 2023
CO-OP
- ▪built NLP-driven product search with BERT embeddings + XGBoost re-ranking. sub-100ms latency end to end
- ▪shipped prompt-based extraction over OCR'd PDFs at ~87% recall, cutting manual data entry by 60%+
scroll
JulAugSepOctNovDecJanFebMarAprMayJunJul
670 CONTRIBUTIONS IN THE LAST YEAR
LessMore
PROJECTS
hermes
ONLINEcontext memory layer for AI coding agents
- ▪built on 5 research papers
- ▪LoCoEval: 96 F1
- ▪drop-in context management for long sessions
TypeScriptBunGeminiContext Engineering
hermes — session active
▸ compaction detected
loading memories...
✓ auth flow (2h ago)
✓ api schema (4h ago)
✓ test patterns (1d ago)
context restored in 87ms
MORE PROJECTS
mixdora
two-stage fine-tuning pipeline for Mixtral-8x22B (135B)
- ▪SFT on OpenHermes + DPO on UltraFeedback
- ▪DoRA adapters, token-packed training
PythonPyTorchDoRAPEFTTRL
ONLINEmarkdown to styled PDF with MCP server
- ▪instant PDF generation from markdown
- ▪AI agent integration via MCP
TypeScriptMCPPDFMarkdown
ONLINELET'S TALK
whether it's a startup, a research idea, or a side project that needs momentum, reach out.