Applied ML Research
Domain-Specialized SLM Fine-Tuning — Llama 3.2 3B QLoRA
for US Community-Healthcare Q&A
Built an end-to-end pipeline to ingest ~3,200 publicly-available Community Health Needs Assessment (CHNA) reports from US hospitals across all 50 states and territories, then fine-tuned Llama 3.2 3B Instruct into a CHNA-domain Q&A specialist answering complex questions about healthcare infrastructure, resource gaps, and demographic needs.
- Training architecture. LoRA rank 32 across attention + MLP; 4-bit NF4 + double-quant; paged AdamW 8-bit; gradient checkpointing; NEFTune embedding noise; single consumer GPU (RTX 4090, 24 GB). Synthetic Q&A from a local Ollama teacher (Llama 3.1 8B Q4) with faithfulness filtering.
- Hospital-level holdout splits. Train / val / test partitioned by hospital — a hospital in train cannot leak into eval, preventing memorization of near-identical annual CHNAs.
- Multi-metric evaluation. Perplexity, pairwise judging vs base (Qwen 2.5 7B as judge, different family to reduce shared bias), 1–5 rubric, ROUGE-L, and faithfulness on the held-out test set.
Experience
Lead architect of AssessmentX AI — a FastAPI companion service to AssessmentX's Spring-Cloud microservices acting as a one-stop adapter for ingestion, retrieval, search, and RAG-augmented context-chat. Consumed via APIs by AssessmentX; shippable standalone with its own Angular UI.
- Agentic SDK. Domain-agnostic, LangGraph-based SDK so new agents ship by writing only a skill manifest and tools implementation — 5 graph templates, three-tier model resolution, Send-based parallel orchestration, Pydantic structured outputs, and bounded retry reused across agents. Three production specialist agents shipped.
- Document IR ingestion, hybrid retrieval & NL→MongoDB. Type-agnostic Intermediate Representation that dual-indexes all content into MongoDB (structured) and Atlas Vector Search (semantic). Hybrid search via vector + BM25/TF-IDF + RRF + BGE re-rank; provider-agnostic across Cisco BridgeIT.
- Debugger Studio. LangSmith-grade, self-hosted observability for LangGraph agents — HIPAA-aware, stays inside Cisco's network. Live DAG, span tree, latency waterfall, token/cost accounting, LLM-as-judge scoring, and in-browser prompt playground with from-run replay.
- Governed LLM codegen. LLM authors per-pack Python from a markdown spec, gated by AST allowlist → human approval (SHA-256) → frozen-pack snapshot; executes in a subprocess sandbox with memory and time caps.
- Agentic SDLC framework. Multi-IDE toolkit distributed by git subtree as ~55 skills, an age-encrypted secret vault, a TypeScript MCP server (31 tools / 10 resources / 5 prompts), and security gates on CSDL, OWASP ZAP, Trivy, Semgrep, Anchore, and Cisco AI Defense MCP scanning.
Built Cisco's internal assessment-automation platform AssessmentX from the ground up with three other engineers, supporting the CX Healthcare PS-delivery team running HIMSS healthcare assessments.
- Platform build. Java 21 / Spring Boot 3.5 / Spring Cloud Maven monorepo of 10 microservices + Python AI companion (Flask → FastAPI) + Angular 18 standalone UI, wired by Feign over Netflix Eureka; multi-stage Docker on hardened Alpine JRE 21 to Cisco ECR.
- Assessment suite. Led 4-lane parallel delivery of a new assessment type end-to-end, then rationalized into a shared questionnaire component and score model — now the template every new assessment type follows.
- Security hardening. Consolidated secrets into AWS Secrets Manager; drove CSDL and SonarQube compliance — Spring/security/bouncy-castle upgrades, OWASP ZAP, OAuth + CSP hardening.
Customer-facing engineering-consulting role on observability strategy for enterprise customers: Tech Mahindra, Fastenal, Carnival, RNDC.
- Technical SME. Embedded with enterprise AppDynamics customers as architecture-level SME. Codified KPIs with engineering leadership, engaged C-suite to capture the observability vision, translated it into engineering requirements through delivery. Earned top-quartile bonus and a formal recommendation from the engineering lead.
Engineering lead on the Collaboration Workflow Manager (CWM) — Cisco CDCoE's enterprise platform for provisioning users onto Cisco collaboration infrastructure for Bank of America and Apple. Promoted SCE III → IV in Apr 2021.
- CWM architecture lead. Owned data-extraction and data-aggregation microservices; drove cluster consolidation and data-migration services. Led customer rollouts for Bank of America and Apple.
- Webex Contact Center dashboard. Designed and shipped end-to-end independently (Angular + Spring Cloud + MongoDB) — live agent, queue, and contact metrics for Webex CC customers.
- Built Java + jBPM workflow scripts and Spring REST + SOAP-WS APIs for JPMorgan Chase's enterprise provisioning platform. Processed 500K+ requests across a 5-year engagement. Promoted twice; relocated to the US in 2017.
- Continued the JPMorgan Chase engagement from the US delivery center; owned 100% of usability improvements; delivered a complete UI redesign and the platform's log-management system.
Technical Skills
AI / ML Skills
AI / ML Tools
Backend Engineering & Databases
Frontend & Security
Infrastructure & Tools
Education
M.S. Computer Science — Artificial Intelligence
Georgia Institute of Technology · Atlanta, GA (Online)
B.Tech. Electronics & Communications Engineering
Punjabi University, Patiala · Punjab, India