Akshat Bokdia

AI Engineering Intern · Boulder, CO

• Built the response-side compliance classifier for an advisor-facing chat agent — a fine-tuned ModernBERT that scores every generated response against 26 regulatory categories in roughly 37 ms, trained on a 10,000-record corpus of synthetic scenarios and PII-anonymized production traces.
• Built the question-side screen as a pair of models: a LoRA fine-tuned general-purpose guardrail adapted to domain risks its broader taxonomy could not isolate, and a capability classifier reaching 0.97 micro-F1 at 96.7% precision on held-out production traffic.
• Owned the evaluation and serving path end to end — prevalence-adjusted test sets, threshold calibration, real-versus-synthetic slices, and GPU deployment behind authenticated inference endpoints.

PyTorch · Hugging Face · PEFT/LoRA · Modal · vLLM · Logfire

Data Scientist Intern · Chennai, India

• Built a semi-supervised anomaly detection pipeline for time-series network KPIs, combining decomposition techniques with tree-based and density-based models.
• Applied causal inference to estimate the effect of network configuration parameters on user throughput, separating genuine causal impact from confounded correlation.
• Curated a 10,000-sample question-answering dataset and fine-tuned DeepSeek-R1 distilled models with instruction-tuned prompts for domain adaptation.

Python · scikit-learn · statsmodels · DoWhy · Hugging Face · Amazon SageMaker

Data Analyst Intern · Chennai, India

• Created an interactive Power BI dashboard analyzing airline operations and passenger behavior.
• Built a CRM app in Power Apps to manage customer data, track leads, and automate follow-ups.

Excel · Power BI · Power Apps

ML Project Intern · Chennai, India

• Fine-tuned a pre-trained general embedding model for movie review classification, improving retrieval relevance for the team's recommendation surface.
• Preprocessed a 100,000-review dataset and trained a BERT-based summary classifier for Flixjini, the company's streaming search product.

Python · SQL · Sentence Transformers

Graduate Research Programmer · Philadelphia, USA

• Designed an experimental framework to trace multi-step reasoning dynamics through line-level entropy trajectories from next-token probability distributions.
• Built an LLM error detector for OfficeBench agent traces, comparing detection and localization approaches across context windows.

Python · OpenAI SDK · vLLM

Graduate Research Assistant · Philadelphia, USA

• Built a WSI registration, tissue-segmentation, and patch-extraction pipeline to assess whether aged or cross-scanner slides remain usable for downstream training.

Python · OpenCV · SimpleITK · PyTorch

Graduate Research Assistant · Philadelphia, USA

• Built a structured expert database by profiling, cleaning, and standardizing enterprise AI governance practitioners and third-party experts for research and industry collaboration.

Python · Pandas

Summer Research Intern · New Haven, CT (Remote)

• Trained convolutional and autoencoder models to classify structural variants in genomic sequences.
• Built the preprocessing and sequence-embedding pipeline the classifiers were trained on.

Python · TensorFlow · Hugging Face

Summer Research Intern · Baltimore, MD (Remote)

• Modelled a DeepSet architecture to predict power spectra from galaxy catalogues in the CAMELS dataset.

Python · PyTorch

Research Intern · Cincinnati, OH (Remote)

• Developed hybrid feature extraction techniques for brain tumour classification in medical imaging.
• Built deep learning models that improved context-based retrieval accuracy over a brain tumour image corpus — work that became a published conference paper.

Python · TensorFlow · OpenCV