Kinjal Goswami builds
AI systems in public,
mistakes included.

Most teams shipping AI chatbots have no visibility into what the model is actually saying to users once it's live. SentinelAI watches for that — in real time.

A three-signal detection engine catches failure modes a single metric would miss: TF-IDF cosine similarity for semantic drift, Isolation Forest for anomaly detection, and a CUSUM statistical control algorithm for gradual quality decay. A multi-tenant FastAPI backend learns each tenant's "normal" from live usage rather than hardcoding it, with a 50-sample warmup phase to cut false positives during cold starts. Alerts reach Slack within 60 seconds of detection.

Contracts are dense on purpose. This tool fine-tunes DistilBERT on the CUAD dataset (Contract Understanding Atticus Dataset) to flag and classify clauses across 34 legal risk categories — then hands each flagged clause to Llama 3.3 70B via Groq to explain, in plain English, what it actually commits you to.

Upload a contract, get clause-level risk flags with explanations in one pass — a multi-step NLP pipeline collapsed into a single usable tool.

Fine-tuned a pretrained ResNet18 on the EuroSAT Sentinel-2 dataset, swapping the final layer to classify satellite imagery into 10 land-cover classes — forest, water, industrial, residential, and more.

Built a geospatial preprocessing pipeline with rasterio and NumPy to convert raw .tif/.jpg satellite files into normalized tensors the model could actually use, then added an interactive anomaly filter using NumPy matrix masking that strips out vegetation and water classes in real time — leaving just exposed ground and infrastructure.