CV
B.Tech in Computer Science & Engineering (Data Science) | Mechanistic Interpretability · Machine Unlearning · Model Adaptation · Applied ML
General Information
| Full Name | Hetansh Waghela |
| Location | Mumbai, India |
| hetanshwaghela06@gmail.com | |
| GitHub | HetanshWaghela |
| hetanshwaghela |
Education
-
2024 - 2028 B.Tech in Computer Science and Engineering (Data Science)
Dwarkadas J. Sanghvi College of Engineering, Mumbai - CGPA 9.56 / 10
Experience
-
Jul 2026 - Present Forward Deployed Engineer Intern
X-ARC, Remote - Translate ambiguous operational requirements into production AI workflows, contributing across rapid prototyping, system design, integrations, evaluation, and deployment.
- Build and harden agentic systems involving tool use, retrieval, memory, and multi-step orchestration, with emphasis on reliability, guardrails, observability, and real-world feedback.
-
Jun 2026 - Present Research Intern — Glacier Flow Research
Georgia Institute of Technology, Remote (under Prof. Jing-Yung Chen) - Conducting computational research on glacier-flow dynamics through reproducible data-analysis and modelling workflows.
- Reviewing literature, analysing modelling assumptions and failure cases, and documenting experimental methods and negative results for collaborative research.
-
Mar 2026 - Jun 2026 AI Engineer Intern
AhumLabs, Mumbai (Hybrid) - Built RAG for a voice and chat assistant on a multi-tenant trading platform, retrieving tenant knowledge and transaction history for natural-language form filling and vessel and pricing queries.
- Shipped admin tooling for tenant onboarding, form customisation, and knowledge uploads, connecting client-specific fields and retrieval context directly to the assistant's tool schemas and prompts.
Technical Skills
-
Research
- Experimental Design
- Causal Interventions
- Statistical Controls
- Error Analysis
- Reproducibility
-
Interpretability / LLMs
- Sparse Autoencoders
- Activation Analysis
- Probing & Patching
- Machine Unlearning
- LoRA & Reward Modelling
-
Tools
- Python
- PyTorch
- Transformers
- TransformerLens
- SAELens
- JAX / Flax
- Tunix
- NumPy
- Pandas
- Git
Research Projects
-
2026 Cross-Model SAE Feature Transfer
Python, PyTorch, SAEs, Causal Steering - Built a 6,144-prompt evaluation bank and matched 4,000 SAE latents per model using Spearman activation signatures, permutation nulls, subsample-stability filtering, and length/domain residualization, yielding 1,040 Level-2 matches.
- Diagnosed an apparent cross-family null result as a methodology artifact caused by hidden-state scale mismatch and introduced per-model α calibration, recovering a 5.7× effect-size gain (d: 0.04 → 0.23).
-
2025 Ghosts in the Model — Machine Unlearning & Interpretability
PyTorch, Transformers, GPT-2, LoRA - Evaluated Gradient Ascent and NegLoRA across 59 landmark-to-city forget prompts spanning 20 cities, reducing forget-set probability to 7.6% and 0.8%, respectively, with three-seed robustness checks.
- Combined layer-wise linear probes with "Lazarus" activation patching to recover 89.7–100% of forgotten behaviour, showing that behavioural forgetting can suppress rather than erase internal knowledge.
-
2026 Subliminal Unlearning — Controlled Transfer & Mechanism Audits
PyTorch, Activation Analysis, Statistical Testing - Designed a checksum-sealed 504-fact experiment with exposure-balanced pilots and held-out gates; a forensic audit caught a target-exposure confound before it contaminated claims.
- A rank-1 mechanism explained 86.9% of paired activation energy yet failed output reconstruction by roughly 350×, decisively rejecting the hypothesis.
-
2026 Teaching LLMs to Reason — Multi-Domain GRPO Fine-Tuning
JAX, Flax, Tunix, Gemma 2 - Implemented GRPO from scratch across 5 domains, including rollouts, group-relative advantages, KL-regularized updates, and AST-based validators; improved exact accuracy from 16.5% to 60.8% with 100% format compliance.
Engineering Projects
-
2026 ANVIL — Compliance-Grade RAG Over ASME Standards
Python, Pydantic, NetworkX, sqlite-vec, FastAPI - Built a compliance-grade RAG system for pressure-vessel queries with auditable provenance, achieving 0.950 pass rate and 1.000 calculation correctness on a 100-example SPES-1 benchmark with NVIDIA NIM.
- Engineered a hybrid retrieval stack combining BM25, sqlite-vec embeddings, and a typed NetworkX knowledge graph via RRF and reranking, reaching 1.000 retrieval recall@10.
-
2026 ImmunoPath — Immunotherapy Decision Support Pipeline
Python, PyTorch, LoRA, Hugging Face, Gradio - Built a 4-model clinical decision-support pipeline combining MedGemma, TxGemma, Path Foundation, and MedSigLIP to predict 8 immune biomarkers directly from routine H&E slides.
- Fine-tuned a 4B vision-language model with LoRA on 7,574 diversity-selected patches from 950 TCGA patients; achieved 100% JSON/schema compliance and TIL MAE of 0.16 on 94 held-out patients.
-
2025 MythForge — Agentic Mythology Engine
Python, Gemini, LlamaIndex, Neo4j, Mem0 - Architected a multi-phase generative pipeline transforming a single seed concept into a fully interconnected mythology using Gemini, LlamaIndex RAG, and Pydantic-validated entities.
- Built an MCP server with 7 tools enabling external AI agents to create, query, and expand mythologies programmatically; placed 4th at the MCP 1st Birthday Hackathon.
Open Source Contributions
- Contributor to Stanford CRFM HELM, supporting reproducible LLM evaluation.
- Contributor to TransformerLens, supporting mechanistic interpretability tooling.
Achievements
- 4th place (Creative, Track 2) at Anthropic & Gradio's MCP 1st Birthday Hackathon with 7,000+ global participants.
- Second Runner-Up at TechnoGrad 3.0 (DJSCE, Mumbai).
- Top 40 rank in DataHack 3.0, Mumbai's largest datathon.
Extracurriculars
-
Sept 2025 - May 2026 Co-Committee Member (Machine Learning)
DJS S4DS, Dwarkadas J. Sanghvi College of Engineering, Mumbai - Designed ML/NLP problem statements for hackathons and workshops; built dataset curation and automated evaluation pipelines.