---
title: "Tanuj Dargan — ML systems and interpretability"
description: "I ship production ML systems and study what language models represent inside. CS at the University of Victoria, graduating April 2027."
canonical: https://tanujdargan.com/
resume: on request
updated: 2026-11-07
---

# Tanuj Dargan

> I ship production ML systems and study what language models represent inside. CS at the University of Victoria, graduating April 2027.

Based in Toronto and Victoria. Resume (PDF, Aug 19 2026, source of truth): on request (https://tanujdargan.com/#resume or dargantanuj@gmail.com)

## At a glance

- **Amazon Robotics**: LLM planning assistant in production across 180+ fulfillment sites. Return offer.
- **UVic CfAR**: Led onboard-computer firmware for MarmotSat, now in orbit after a Falcon 9 launch.
- **Apart Research**: Showed SFT loss masking decides whether a model voices shutdown distress.

## Selected work

### Loss masking decides whether a model voices shutdown distress

- Where: Apart Research, Digital Minds sprint
- When: Aug 2026
- Result: Output-based welfare signals are unreliable
- Finding: Assistant-only loss maximizes expressed distress; any user-turn loss restores a neutral register with no capability loss. Calm-sounding models still represent shutdown negatively.
- Write-up: https://tanujdargan.com/research/loss-masking-distress.md (web: https://tanujdargan.com/research/loss-masking-distress)

### Steering a language model along emotion directions

- Where: Method replicated on Qwen2.5-1.5B-Instruct for this site
- When: 2026
- Result: A sharp flip, not a gradient
- Finding: Joy steering is a switch, not a dial: between +0.3 and +0.6‖h‖ the read-back projection jumps from −8.0 to +9.6.
- Note: these numbers come from a replication run for this site, not from his main project.
- Write-up: https://tanujdargan.com/research/activation-steering.md (web: https://tanujdargan.com/research/activation-steering)

### An LLM planning assistant in production at Amazon

- Where: Amazon Robotics, SDE intern
- When: May–Aug 2026
- Result: In production, with a return offer
- Finding: Built into Optimus, the labor-planning platform for 180+ fulfillment sites in North America and Europe. RAG over historical labor plans and site configs.
- Metrics: 180+ fulfillment sites in NA and EU; 60%+ cut in planner time-to-insight; $136.2M projected yearly savings

### Satellite firmware, now in orbit

- Where: UVic Centre for Aerospace Research
- When: 2025–26
- Result: Launched July 2026, NORAD 69912
- Finding: Led MarmotSat's onboard-computer firmware in C and the TTC board design. A CI and hardware-in-the-loop pipeline cut testing from 180 minutes to 15.

### Deep-water temperature from satellite imagery

- Where: SOLIDS Lab, UVic
- When: 2025
- Result: R² 0.936 against deep-water sensors
- Finding: A multimodal model on Ocean Networks Canada sensors and satellite imagery, validated with the National Research Council Canada.
- Note: the ocean demo on the site is illustrative; only R² 0.936 is a measured result.
- Write-up: https://tanujdargan.com/research/ocean-sensing.md (web: https://tanujdargan.com/research/ocean-sensing)

## Experience

### Software Development Engineer Intern, Amazon Robotics

- When: May – Aug 2026
- Where: Toronto
- Summary: Shipped an LLM planning assistant to production in Optimus, the labor-optimization platform for 180+ fulfillment sites. Return offer.
- LLM-powered planning assistant in production across 180+ fulfillment sites in NA and EU: natural-language plan queries, anomaly detection and labor recommendations. Projected to save $136.2M a year.
- RAG pipeline over historical labor plans and site configurations. Cut planner time-to-insight by 60%+.
- Multi-turn interface on the optimization engine for what-if scenarios, returning re-optimized plans without full solver re-runs.

### AI/ML research mentee, AI Mentorship Program, RBC Borealis

- When: Mar – May 2026
- Where: Remote
- Summary: ML pipeline flagging at-risk youth with chronic illness from 1,500+ longitudinal engagement records.
- PyTorch models on 1,500+ longitudinal records (3+ years, 20+ behavioural features). UMAP clustering isolated 4+ user cohorts.
- Analytics dashboard for 4+ clinical and program teams: retention curves, engagement trends, disengagement risk flags.

### Software developer & undergraduate researcher, UVic Centre for Aerospace Research

- When: Apr 2025 – Apr 2026
- Where: Victoria
- Summary: Led OBC firmware and TTC board design for MarmotSat, now in orbit after launching on a SpaceX Falcon 9.
- Led onboard-computer firmware (C), TTC board design and hardware/software integration for MarmotSat.
- CI/HIL pipeline with GitHub Actions and SWD flashing, 80+ checksum rules per commit. Testing went from 180 min to 15.
- C/Java YAMCS integration that removed 20+ hours of manual work per release.

### Undergraduate researcher · Prof. Sean Chester, DEIA Lab, UVic

- When: May 2025 – Feb 2026
- Where: Victoria
- Summary: Custom retrieval on a modified cuVS fork: 1–4M QPS at billion-vector scale, 90–95% recall, under 250 ms.
- CAGRA and IVF-PQ across 8+ NVIDIA GPUs and 2 TB of index.
- Automated stress tests on Kubernetes GPU clusters: 4.3× throughput scaling, NVLink saturation exposed.

### Lead AI developer, Pear Care

- When: Jul 2025 – Apr 2026
- Where: Remote
- Summary: Autonomous medical-AI agents across 1,000+ clinical specialties.
- LangGraph agents with MoE routing and task-introspective reasoning, served via FastAPI on GCP.
- Quantized LoRA/PEFT deployments cut memory 40% at 95%+ F1 on internal triage benchmarks.
- Multi-tenant GCP/AWS infrastructure with Redis caching and DynamoDB, validated through load testing.

### Software engineering fellow, MLH Fellowship · Apache Airflow

- When: May – Aug 2025
- Where: Remote
- Summary: Docker-based IDE debugging for Airflow (50K+ stars): setup-to-debug from hours to minutes.
- Step-through debugging in VS Code and PyCharm with dynamic port exposure. Team of 3 with Royal Bank of Canada engineers.

## Education

- University of Victoria: B.S. Computer Science (accelerated, 3 years). Graduating April 2027. International Scholarship.
- Relevant coursework: Quantum Computing, Computer Architecture, Numerical Analysis, Communications & Networks

## More work

- **Billion-vector search on GPUs**: DEIA Lab, Prof. Sean Chester. A modified cuVS fork: 1–4M QPS at 90–95% recall, under 250 ms. https://tanujdargan.com/research/billion-vector-retrieval.md
- **Fine-tuning Qwen2.5-1.5B without gradients**: Evolution strategies on a LoRA with a binary GSM8K reward. Adapter on Hugging Face. https://tanujdargan.com/research/es-lora.md
- **Optimate, Hack the North 2025**: Sep 2025. AI underwriting dashboard (Next.js, Cohere LLMs, AWS DynamoDB). Won 2 prizes out of 300+ teams: YC Unicorn Prize and Federato Best RiskOps. The Federato CEO validated 40% faster review cycles and 15% fewer underwriting errors versus manual workflows. https://github.com/jruttan1/Optimate
- **Cut the Wire: collusion-probe ablations under controls**: Apart Research, AI Collusion sprint, Oct 2026. In Signal Table, a steganographic betting game, most of the large effect of ablating collusion-probe directions decomposes into format damage and generic cost; a modest 8–15 point late-layer effect remains in Qwen2.5-7B. NARCBench's steganography tier is separable by context length alone (AUROC 1.000). https://apartresearch.com/sprints/projects/cut-the-wire-collusionprobe-ablation-effects-shrink-or-relocate-under-format-dose-and-placement-controls-10ps
- **Deleting hazardous knowledge from model weights**: Independent research, Oct 2026. Deleting a 16-direction subspace from every layer of Gemma-2-2B keeps 72% of its WMDP-cyber forgetting against a LoRA relearning attack vs 13% for RMU and NPO (p = 0.0008); against full fine-tuning the edge disappears.
- **AIMO Progress Prize 2**: Kaggle, Nov 2024 – Mar 2025. Silver medal, top 10% of 2,200 teams, 54% accuracy (finishing score 27). Multi-stage LLM pipelines with fine-tuned open-weight models, NeMo-Skills reasoning strategies and in-context-learning ablations.
- **Beri**: A food app I'm building for iOS. https://getberi.com

## Research write-ups

- [Steering models along emotion and persona directions](https://tanujdargan.com/research/activation-steering.md): Emotion and persona vectors, extraction to steering (2026)
- [Does loss masking shape how a model talks about being shut down?](https://tanujdargan.com/research/loss-masking-distress.md): SFT loss-masking regimes vs shutdown distress (Aug 2026)
- [Fine-tuning Qwen2.5-1.5B without gradients](https://tanujdargan.com/research/es-lora.md): LoRA fine-tuning with evolutionary strategies (May 2026)
- [Keeping billion-vector search fast on GPUs](https://tanujdargan.com/research/billion-vector-retrieval.md): cuVS retrieval at 1–4M QPS, billion-vector scale (May 2025 – Feb 2026)
- [Predicting deep-water temperature from satellites](https://tanujdargan.com/research/ocean-sensing.md): Ocean-sensing models with NRC Canada (2025)

## Contact and links

- Email: dargantanuj@gmail.com
- Resume (PDF, Aug 19 2026): on request (https://tanujdargan.com/#resume or dargantanuj@gmail.com)
- GitHub: https://github.com/tanujdargan
- LinkedIn: https://www.linkedin.com/in/tanujdargan
- X: https://x.com/DarganTanuj
- Website: https://tanujdargan.com/

---

Machine-readable index: https://tanujdargan.com/llms.txt · Everything in one file: https://tanujdargan.com/llms-full.txt