Anmol Kumar

A reusable Python framework for retrieval-augmented AI features - confidence-scored, hallucination-aware, and context-managed so teams don't have to reinvent the plumbing.

• Designed a pluggable retrieval core with ranked multi-source fusion, conversational context management, and per-claim confidence scoring - so the framework can say 'I don't actually know' instead of guessing.
• A hallucination detection layer that flags weakly-grounded responses before they reach a user, which matters a lot when the user is a support engineer looking at a customer ticket.
• Adopted across 4+ product teams inside BMC - which pulled new AI POCs from ~3 weeks of scaffolding down to under 5 days.
• Quietly became the standard pattern for how retrieval-augmented features get built across the IMS, Db2, and z/OS product lines.

A tool-calling multi-agent system that takes the first pass at root-cause analysis on IMS mainframe telemetry, so SMEs walk into a review, not a blank page.

• A set of specialist agents - log triage, metric correlation, knowledge retrieval, narrative summarization - with a supervisor agent that routes work based on the shape of the incident.
• The tool layer wraps live IMS, Db2, and z/OS data, so agents pull fresh telemetry at decision time instead of guessing from whatever was in the context window.
• On supported incident classes, SME investigation time dropped by an estimated 60-70% - from 2-4 hours of manual digging to a few minutes of reviewing what the agents already surfaced.
• Part of the work recognized with the BMC IMS Impact Makers Award (Q1 FY26).

Real-time log-stream analysis that turns mainframe chatter into anomalies and early-warning signals you can actually act on.

• Pulls in real-time log streams from IMS, Db2, and z/OS and classifies events against a domain-specific taxonomy we keep inside the RAG framework.
• Surfaces predictive signals - capacity cliffs, recurring subsystem issues, subtle regressions - rather than only after-the-fact alerts.
• Fed directly into 3 product roadmap decisions by making previously invisible patterns legible to product managers and SMEs.

A conversational copilot that replaces manual dashboard-hopping with natural-language questions against live enterprise data.

• Natural-language access to system observability data across IMS and database performance products - support engineers stop hunting across dashboards and just ask the copilot.
• Live data retrieval through the agent tool layer, so answers reflect the system as it is right now rather than a stale snapshot.
• Built on the internal RAG framework, with confidence scoring so the copilot can gracefully say 'I'm not sure' when grounding is weak - better than a confident wrong answer.

Supervised fine-tuning of sub-3B transformer models to show on-prem AI is realistic for air-gapped mainframe shops that can't touch hosted APIs.

• Fine-tuned small transformer models on modest hardware, with domain-specific corpora drawn from IMS and Db2 documentation and real telemetry patterns.
• Showed that sub-3B parameter models can handle the most common domain-specific tasks at a small fraction of what frontier models cost per inference.
• Unblocked AI conversations with air-gapped enterprise customers who simply can't send data to hosted LLM APIs - suddenly there was a believable story for them.

A unified, AI-powered healthcare platform I built to pull hospitals, clinics, labs, and patients into a single secure place - so medical records stop living in silos.

• A document extraction pipeline that takes medical report PDFs and images through OCR and turns them into structured parameter tables.
• A clinical analysis layer on AWS SageMaker-hosted Med42 (Llama3-8B) that generates severity-aware summaries and patient-friendly explanations on top of those parameters.
• A conversational health assistant with role-isolated sessions for patient, doctor, and admin dashboards - and a context strategy that prioritizes structured parameters over raw OCR text inside the 2048-token window.
• Trend tracking across reports with Recharts visualisations, so patients and doctors can see parameters move over time instead of reading isolated one-off results.