AI researcher and engineer. I measure whether a model’s explanation is actually true.
A model’s explanation for its own output is often made up after the fact.
I reverse-engineer transformers with causal circuit analysis (activation patching, ablation, sufficiency and comprehensiveness scoring against ERASER metrics) to test whether a stated explanation is causally load-bearing, not just fluent.
Creator & Maintainer
Glassbox AI, Trier
Feb 2026 – Present
Built and maintain glassbox-mech-interp (PyPI, MIT core): causal circuit discovery via activation patching, with an ACDC path for exact edge-level discovery, extended to Llama-3, Mistral, Gemma and Phi-3 through a dedicated GQA/RMSNorm architecture adapter. Automates EU AI Act Annex IV documentation generation from circuit analysis into structured JSON. Shipped with a CLI, automated test suite, GitHub Actions CI, and a live Hugging Face demo — 15–37x faster than ACDC on GPT-2 Small (single-run, Apple M2 Pro).
Machine Learning Engineer
BaseMotion AI, Berlin
Nov 2025 – Apr 2026
Applied mechanistic circuit analysis (activation patching, ablation) to GPT-2 Small, scoring explanations against ERASER sufficiency/comprehensiveness metrics. Built the pipeline in PyTorch and TransformerLens with an automated test suite and CI for reproducible runs. Result: r = 0.009 between model confidence and explanation faithfulness: confidence carries no signal. Separately measured that fluent LLM-written explanations score 99% on perceived quality (ERASER), up from 60%, with no corresponding gain in mechanistic faithfulness.
Independent Researcher, Project Thesis & AI Safety
Hochschule Trier
Apr 2024 – Sep 2025
Synthesized 25+ papers across AI alignment, scalable oversight, and interpretability into structured technical reports, benchmarking robustness methodologies and cataloguing frontier LLM failure modes to establish baseline quantitative metrics for the field.
AI Engineer
ICEICO Technologies, Nagpur
Feb 2021 – Mar 2023
Built PyTorch training pipelines and tuned hyperparameters for internal prediction models, establishing reproducible baselines for testing. Shipped a Retrieval-Augmented Generation pipeline with LangChain to automate knowledge retrieval across production datasets.
Glassbox AI
Feb 2026 – Mar 2026
An open-source mechanistic interpretability toolkit that finds the causal circuit behind a transformer prediction and turns it into EU AI Act Annex IV documentation automatically. 4,400+ PyPI downloads to date, live HuggingFace demo, CI passing on 9 architecture families.
View on GitHub →BFSI Credit Intelligence
May 2026
Agentic loan-underwriting platform for Indian BFSI: six LangGraph agents (document validation, financial analysis, XGBoost + SHAP credit scoring, Isolation Forest fraud detection, RBI/DPDP compliance, final decision) fan out in parallel behind five MCP servers, with every decision explained by SHAP and published to Kafka for audit. Turns a 3-5 day manual process into a sub-5-second, regulator-ready decision.
View on GitHub →Azure ML Pipeline
Feb 2026 – Jun 2026
End-to-end MLOps pipeline on Azure ML SDK v2 for heart disease prediction: a four-stage pipeline (data prep, training, evaluation, registration) running on auto-scaling compute clusters, MLflow tracking every run, and automatic promotion to a managed REST endpoint once a model clears the accuracy threshold.
View on GitHub →Azure AI RAG System
Jan 2026 – Jun 2026
Production RAG pipeline on Azure OpenAI and Azure AI Search: documents chunked at 1,000 tokens with 200-token overlap, embedded at 1536 dimensions, retrieved with hybrid vector plus BM25 search, and answered by GPT-4o-mini constrained to the retrieved chunks only. Every response carries a citation back to its source document.
View on GitHub →Causally-Grounded Mechanistic Interpretability
M.Sc. Thesis · Dec 2025 – May 2026
Evaluates a model circuit the same way you'd evaluate an explanation: does removing it break the output (sufficiency), and does keeping everything else still break it (comprehensiveness)? On GPT-2's IOI task the discovered circuit beat an attention-only baseline on 75% of prompts, scored 100% sufficiency, but only 22% comprehensiveness. Sufficiency alone oversells how reliable an explanation is.
View on GitHub →Enterprise Agentic AI Platform
May 2025 – Jun 2025
A multi-agent system on LangGraph coordinating three MCP servers (Postgres, documents, notifications) with hybrid RAG, built to take a rough MVP into something enterprise-shaped. FastAPI streamed data over WebSockets to a Next.js frontend, Celery handled async work, Langfuse and Prometheus/Grafana gave it real observability, all containerized with Docker Compose.
View on GitHub →Two papers live on arXiv. Rest unpublished — links go live on release.
Explanation Multiplicity: Circuit-Level Interpretability Evidence Does Not Survive Defensible Analytic Variation
arXiv:2608.13754 · 2026
Circuits recovered by mechanistic interpretability tools are not unique: the same model, task, and method can return different circuits depending on ablation operator, corruption distribution, and metric choice. This measures how far an EU AI Act Annex IV conformity claim built from a discovered circuit moves across the space of defensible analytic specifications, and tests whether that movement is distinguishable from a size-matched random circuit baseline.
Causal Attribution for Agentic Decisions: Estimators, Coupling, and a Traceability Specification
arXiv:2609.06445 · 2026
Multi-agent systems produce decisions that no single deployer can fully attribute. This shows the standard marginal-effect estimator can return the same value for a causally inert step and a decisive one — under common random numbers, it returns an exact zero for a decisive step on roughly one run in ten, so a zero does not certify that a step did nothing. Derives a coupling that preserves estimability and a traceability specification bridging the gap between the EU AI Act's Article 86 explanation rights (Aug 2026) and its deferred documentation requirements (Dec 2027).
What Monitorability Metrics Miss
Target ICLR 2027 · Sep 2026
The published chain-of-thought monitorability autorater was validated against synthetic degradations, a broken reasoning trace. That checks whether the metric notices a chain of thought falling apart, not whether it notices a well-formed one that conceals its true cause, which is the actual safety concern. This runs the validation the metric's own authors flagged as outstanding.
In progressConfidence doesn’t predict faithfulness.
Model confidence predicts explanation faithfulness almost not at all: r = 0.009, measured with causal circuit analysis on GPT-2 Small (Indirect Object Identification task). The 6-head IOI circuit explains 61.4% of the logit difference and is sufficient on its own, but removing it still leaves 78% of the signal intact through backup heads: it isn’t necessary.
Glassbox AI: audits that write their own paperwork.
Glassbox AI finds the causal circuit behind a transformer prediction with activation patching, scores it for sufficiency and comprehensiveness, and writes EU AI Act Annex IV documentation as structured, content-hashed JSON, about 1.8s per analysis on GPT-2 Small (Apple M2 Pro), 15–37× faster than ACDC depending on circuit size. 943 tests passing in CI across 9 architecture families, GPT-2 through Llama-3.
Find the causal circuit once;
ship EU AI Act documentation with every audit.
Programming
ML & Research
LLM & Agents
Cloud & MLOps
Focus
Languages
Glassbox AI ships with 943 tests passing across 9 architecture families and 4,400+ PyPI downloads to date. Every number here is checked against the repo, not rounded up.
M.Sc. in Interdisciplinary Engineering, AI/ML (Hochschule Trier), researching mechanistic interpretability and LLM faithfulness evaluation. I build evaluation tooling that turns interpretability findings into audit-ready evidence, not slide-deck claims.
Trier, Germany. Available full-time now. German residence permit, no sponsorship required — open to relocation within Germany and the DACH region. Open to full-time ML engineer, AI safety and governance, and PhD positions.
M.Sc. Interdisciplinary Engineering — AI & ML
Hochschule Trier · 2023–2026 · Grade 1.8, Thesis 1.0
B.Tech, Automotive Engineering
Government College of Engineering and Research · 2018–2021 · Grade 8.77/10, Top 5%
- Microsoft Certified: Azure AI Engineer Associate (AI-102), Azure AI Fundamentals (AI-900)
- Anthropic, Claude Code in Action
- BlueDot Impact, Technical AI Safety (30-hour cohort)
- Google, AI Essentials